JP4052609B2 - Pattern recognition apparatus and method - Google Patents

Description

[0001]
BACKGROUND OF THE INVENTION
The present invention relates to an apparatus and method for recognizing a pattern, and more particularly to an apparatus and method for recognizing a pattern that is stable against a reflected luminance pattern subjected to illumination fluctuations.
[0002]
[Prior art]
The basic framework for pattern recognition is to express the characteristics of patterns belonging to the category to be recognized in some form, and collate and evaluate how well the characteristics of unknown input patterns match the characteristics of this category. . Information representing the characteristics of this category is called a dictionary, and a measure indicating the degree of match is called similarity. The potential of the pattern recognition system depends on how these two are defined and how the processing system is implemented. Greatly affects.
[0003]
Here, the framework for pattern recognition is briefly explained.
[0004]
A pattern p composed of N scalar quantities associates each scalar quantity with one axis, and each scalar quantity value is a point on the corresponding axis (or a point from the origin in the N-dimensional superspace S). It can be regarded as a position vector v). In the case of the image pattern p, each pixel value corresponds to this scalar quantity. This is a pattern vectorization method used in many pattern recognition apparatuses. At this time, how similar the two patterns are can be evaluated by calculating the similarity based on the distance between the patterns. This is the simplest pattern recognition framework.
[0005]
There are various ways of defining the similarity between two patterns. For example, the reciprocal of the Euclidean distance between two points, the cosine of each point's position vector (this is called a feature vector) (ie, from one to the other) Projective length). In this case, the greater the similarity, the more similar the two patterns.
[0006]
In general, when a plurality of known patterns belonging to a certain category are given and one unknown pattern is given, in order to evaluate how similar this unknown pattern is to this category, the above-mentioned 2 This can be done by calculating the similarity between patterns.
[0007]
In this case, the similarity between the unknown pattern and all the known patterns is obtained individually, and for example, the largest similarity is set as the similarity with this category. In this method, all known patterns are individually stored as a dictionary representing the characteristics of a category. However, with this method, as the number of known patterns increases, the cost of similarity calculation and the memory space for storing the known patterns increase. In addition, since there are patterns similar to each other in the known patterns, there is a waste of calculating the degree of similarity many times with similar patterns.
[0008]
Therefore, instead of storing the known patterns individually, it is necessary to describe the characteristics of the known pattern set (ie, category) with less information.
[0009]
As a method of expressing the features of a category, when a plurality of k N-dimensional feature vectors {v: v1, ..., vk} (corresponding to the aforementioned known patterns) belonging to a certain category are given, these feature vectors are There is a method in which an M-dimensional superspace L based on M (M ≦ N) orthonormal vectors {e: e1,..., EM} obtained by principal component analysis is used as a dictionary. The M-dimensional superspace L is the N-dimensional superspace S itself (M = N) or a subspace L (M <N) of the number of dimensions M of the N-dimensional superspace S. For convenience, the M-dimensional superspace L Is called a dictionary subspace of dimension M.
[0010]
According to this method, even if the number k of known patterns (also referred to as teaching patterns) is much larger than N, it is only necessary to store at most M (≦ N) basis vectors as a dictionary. Although the degree of similarity with the dictionary subspace L when one unknown feature vector v (corresponding to the above-mentioned unknown pattern) is given has been defined in various ways, the unknown feature vector v is projected onto the dictionary subspace L. The method of defining the length (calculated as the sum of squares of the lengths of the feature vectors v projected to all the base vectors {e: e1, ..., eM} of the subspace L) as the similarity is the subspace It is called the law.
[0011]
Detailed information related to the definition of the dictionary representation and similarity including the subspace method can be found in Ref. [1] (Erki Oya (translated by Hidemitsu Ogawa), “Pattern Recognition and Subspace”, Industrial Books, 1986), and [2] (by Taizo Iijima, “Pattern Recognition Theory”, Morikita Publishing, 1989).
[0012]
The various conventional methods described above describe the dictionary expression as a vector (or a group of vectors) or a subspace, but all of the inputs that are the targets for similarity calculation with the dictionary are vectors (that is, only one The pattern is common.
[0013]
On the other hand, Maeda uses the subspace Ld (similar to the subspace method) as a dictionary and 1 as an unknown input in Ref. [3] (Kenichi Maeda, “Pattern Recognition Method”, Japanese Patent Laid-Open No. 60-57475). Subspace Li (referred to as input subspace) by a plurality of patterns {p: p1, ..., Pk} generated by artificially adding positional deviation to one input pattern p A mutual subspace method is proposed in which the maximum cosine of the dictionary subspace Ld is defined as similarity (referred to as intersubspace similarity).
[0014]
Yamaguchi et al. [4] (Osamu Yamaguchi et al., "Face Recognition System Using Moving Images", IEICE Technical Report PRMU97-50, 1997) generates mutual subspaces Li from input time series patterns. A subspace method is proposed and applied to human face recognition. In this document, it is shown that the mutual subspace method exhibits better recognition performance than the conventional subspace method when the same set of input time series patterns and the same set of dictionary subspaces are given and compared.
[0015]
In particular, in the subspace method, a phenomenon in which the similarity greatly fluctuates due to a change in the face direction or facial expression of the person to be recognized, but it has been confirmed that the mutual subspace method stabilizes at a high value of similarity.
[0016]
A difference between a recognition method such as a subspace method in which a similarity calculation target is one pattern and a mutual subspace method will be schematically described with reference to FIG. 41 in the figure is a dictionary subspace belonging to category A, and 42 in the figure is a dictionary subspace belonging to category B. “×” in the figure is an input pattern belonging to category A, and as shown in the upper part of the figure, pattern 43 is close to category A dictionary subspace 41 and pattern 44 is close to category B dictionary subspace 42. Therefore, in the recognition method in which the similarity calculation target is one pattern, the pattern 43 is classified into the category A and the pattern 44 is classified into the category B. On the other hand, as shown in the lower part of the figure, in the mutual subspace method for generating the input subspace 45 from the input pattern, the input subspace 45 is closest to the dictionary subspace 41 of category A. Are classified into category A.
[0017]
Thus, the similarity of the mutual subspace method is stable against the existence of a variation pattern far from the original category such as the pattern 44, thanks to the existence of a pattern close to the original category such as the pattern 43. I understand.
[0018]
The similarity stability in the mutual subspace method can be qualitatively interpreted as a synergistic effect of the following two stability.
[0019]
(1) Stability of the input subspace: In the mutual subspace method, the input subspace is extended by a large number of feature vectors. Therefore, random minute fluctuations that always appear in many feature vectors, or rare that appear only in certain feature vectors. Variation is less likely to affect the entire input subspace. That is, the stability of the input subspace depends on the magnitude and frequency of the fluctuation.
[0020]
(2) Stability of similarity between subspaces: In the mutual subspace method, the maximum cosine of the input subspace and the dictionary subspace, that is, the cosine of two vectors belonging to both subspaces (which is innumerable) Since the largest value (maximum cosine) is used as the similarity, even if there is some variation in the input subspace, the similarity itself is difficult to vary simply by changing the vector set.
[0021]
Based on the above, we consider the phenomenon in Ref. [4]. A pattern that has been subjected to a random minute change that has not been learned is always input due to a change in face orientation or facial expression, and a pattern that has received a large change that has not been learned (rare change) is rarely input. In the subspace method, the degree of similarity reacts and vibrates immediately even with unlearned minute fluctuations. In addition, in the subspace method, the degree of similarity is unacceptably decreased for a large fluctuation that has not been learned yet.
[0022]
On the other hand, in the mutual subspace method, the input subspace can remain close to the dictionary subspace as long as it is a minute change or a rare change without distinguishing between learned and unlearned changes (stability of the input subspace). . As long as the input subspace stays close to the dictionary subspace, the similarity between subspaces can maintain a high value even if the input subspace fluctuates slightly (stability of similarity between subspaces).
[0023]
Further, in the mutual subspace method, the tolerance to both the magnitude and the frequency of the fluctuation can be simultaneously enhanced by learning the fluctuation in the dictionary subspace. This stability and learning are important properties in providing a pattern recognition system that is robust against fluctuations.
[0024]
In the subspace method (other methods in which one input pattern is used), it is possible to provide resistance against rare fluctuations by post-processing such as averaging the similarity to a plurality of input patterns in a predetermined period. Even so, since the similarity to each input pattern varies, it is difficult to obtain similarity stability corresponding to the synergistic effect of the stability of the input subspace and the stability of the subspace similarity.
[0025]
By the way, an image obtained by photographing an object under a certain illumination environment (that is, an image by reflection luminance) is a solid shape and surface reflectance pattern of the object, illumination spectral intensity (color and brightness), illumination direction and imaging. It depends on the direction. The three-dimensional shape of the object and the surface reflectance pattern are features unique to the object, and the feature appears in the captured image as an image pattern unique to the object (that is, a reflection luminance pattern). This is the basis on which an object can be recognized in an image pattern.
[0026]
However, this image pattern is a deformation of an object (for example, a three-dimensional shape change such as a facial expression change in the face), an illumination variation (for example, a change in illumination spectral intensity or illumination direction due to time zone or weather), and an object posture variation (for example, in the face). Changes in imaging direction such as a change in orientation) are affected in various ways. In particular, regarding the illumination direction, it is difficult to learn patterns of various illumination directions in the dictionary subspace in advance, despite the instability that the shading of the image pattern changes dramatically. Unlearned lighting fluctuations give unlearned luminance fluctuations to all input patterns. For this reason, even the mutual subspace method, which is considered to be resistant to fluctuations, cannot guarantee the resistance to the rare fluctuations and minute fluctuations, which causes a serious decrease in similarity and erroneous recognition.
[0027]
The influence of illumination fluctuation will be schematically described with reference to FIG. 51 in the figure is a dictionary subspace belonging to category A, and 52 in the figure is a dictionary subspace belonging to category B. “×” in the figure is an input pattern belonging to category A, and has moved as a whole due to the influence of illumination fluctuations. As shown in the upper part of the figure, among such illumination variations, pattern 53 is close to category A dictionary subspace 51 and pattern 54 is close to category B dictionary subspace 52. However, the number of patterns close to category A has decreased due to movement of the entire pattern. At this time, as shown in the lower part of the figure, since the input subspace 55 generated from the input pattern is halfway between the category A dictionary subspace 51 and the category B dictionary subspace 52, the similarity is The input subspace 55 may even be classified into category B depending on the input pattern.
[0028]
In order to prevent such a decrease in similarity and erroneous recognition due to illumination fluctuations, it is necessary to perform recognition using feature vectors that are not easily affected by illumination and that preserve object-specific features. Since the illumination variation gives a strong variation to the reflected luminance pattern of the object, the influence of the illumination variation is inevitable as long as the reflected luminance pattern itself is a feature vector. On the other hand, it is known that certain feature vectors derived from the reflected luminance pattern have constancy that is hardly affected by illumination conditions.
[0029]
For example, in reference [5] (Yael Moses et al., “Face Recognition: the Problem of Compensation for Changes in Illumination Direction”, Proceeding of ECCV'94, 1994), reflected luminance is an effective feature against illumination variation. A luminance gradient image obtained by differentiating an image (hereinafter referred to as a differential image) has been proposed.
[0030]
The first-order differentiation to the reflected luminance image corresponds to an operation for obtaining the luminance change rate, and a portion (edge) where the luminance changes stepwise is extracted. The second-order differentiation to the reflected luminance image corresponds to an operation for obtaining the rate of change of the reflected luminance change rate, and a portion where the luminance changes in a ridge shape or a valley shape is extracted.
[0031]
The surface of the object has high and low reflectance patterns peculiar to the object, and the low reflection luminance always changes regardless of the illumination intensity at locations where the reflectance is low, and the reflection luminance changes depending on the illumination intensity at locations where the reflectance is high. To do. For this reason, the first-order differential image represents (a) a change in reflectance at a location where illumination is applied and (b) a change in shadow at a location where high reflectance is applied. (A) is an edge where there is no fluctuation in position even if there is an intensity fluctuation due to illumination fluctuation unless the attitude of the object changes, and (b) is an intensity and position which both fluctuate due to illumination fluctuation even if the attitude of the object does not change. It is an edge.
[0032]
The first effect of the differential image with respect to illumination variation is that (a) preserves the characteristic of the reflectance pattern unique to the object according to the posture of the object. This shows the validity of using the differential pattern for object recognition. The second effect of the differential image on the illumination variation is that (a) appears stronger than (b) as the height difference of the reflectance pattern of the object surface increases, that is, the texture of the object is stronger. In particular, when the object has a smooth curved surface, the shadow due to illumination shows a gentle change, so that (b) is further weakened. This means that unlearned illumination fluctuation, which was a large-scale fluctuation in the reflected luminance image, can be suppressed to an unlearned fluctuation relating to (b) in the differential image (localization and miniaturization). The differential image is considered to have illumination variation constancy due to such an effect. On the other hand, however, the presence of (b) does not completely eliminate the influence of illumination fluctuations. Therefore, it can be seen that it is necessary to use in combination with a recognition method capable of absorbing local minute fluctuations in (b) in order to give illumination fluctuation constancy to the pattern recognition system.
[0033]
In fact, in the experiment according to the literature [5], the inter-pattern distance was used as the similarity, so that the improvement of the recognizability by differentiation was not confirmed. This knowledge means that even if the influence of illumination fluctuation can be suppressed to a local minute fluctuation by devising the feature amount, it is desirable if the most stable pattern recognition method is not selected for this minute fluctuation. The effect is not obtained. Here, the advantage of using the mutual subspace method that has not been suggested heretofore appears.
[0034]
The effect of combining the mutual subspace method and the localization / miniaturization of illumination variation will be schematically described with reference to FIG. 61 in the figure is a dictionary subspace belonging to category A, and 62 in the figure is a dictionary subspace belonging to category B. “×” in the figure is an input pattern belonging to the category A, which is farther than the case where there is no illumination variation, but is approaching the dictionary subspace 61 of the original category A due to the localization / miniaturization of the illumination variation. . As shown in the upper part of the figure, the pattern 63 is close to the category A dictionary subspace 61, and the pattern 64 is close to the category B dictionary subspace 62.
[0035]
However, the number of patterns close to category A is increasing due to movement of the entire pattern due to localization / miniaturization of illumination variation. At this time, as shown in the lower part of the figure, the input subspace 65 generated from the input pattern is closest to the dictionary subspace 61 of category A. The input pattern is safely classified into category A. On the other hand, in the method of calculating the similarity as one pattern, the number of correctly classified patterns is increased or decreased regardless of whether there is illumination variation or localization / miniaturization. A dramatic effect cannot be expected.
[0036]
For an object whose direction and expression change frequently, such as a live human face, the edge (a) of the low-reflectance part such as the eyebrows and eyes moves according to the face direction and expression, and the shadowed edge (b) is illuminated. If the direction is constant, it hardly moves. In other words, if it is assumed that the face orientation and facial expression change frequently (photographs a live human), the movement is different between (a) and (b).
[0037]
Here, considering the eyebrows and eye edges (a), which are facial features learned in the dictionary, (b) is a consistent movement that moves relative to (a) according to the face orientation and the like. It can be regarded as no unlearned minute fluctuation. This is true for face recognition as described in [3] [4], in which various face orientations and expressions were learned under specific limited lighting conditions. Pattern recognition method, which is shown to be superior to the conventional method, and is expected to be highly resistant to minute fluctuations, that is, to use the mutual subspace method with the stability of the input subspace and the similarity between subspaces. A special meaning comes out. However, the differential image application methods that have been proposed in the past did not take any consideration into this point, and as a result, the merit of the differential image could not be fully derived and sufficient performance improvement was not achieved. is there. This is the first problem in the prior art.
[0038]
In an object with a clear reflectance pattern such as a face and a smooth surface shape, the shadow due to illumination appears as a global change in the reflected luminance pattern of the object, and the reflectance change has a finer structure. Appears as a steep luminance change. The above-described differential processing functions as a high-pass filter for the reflected luminance image, and has an effect of removing a fine luminance structure having a high spatial frequency by removing a global luminance change having a low spatial frequency. That is, it can be concluded that high-pass filtering is effective in reducing the influence of illumination fluctuations, as well as the differential processing suggested in the literature [5]. Therefore, by combining the high-pass filtering and the mutual subspace method, the pattern recognition system obtains unprecedented stability for almost all variations including object deformation, object orientation variation, and illumination variation.
[0039]
Another feature amount that utilizes the fact that the influence of illumination appears in the low-frequency component of the reflected luminance image has also been proposed. Reference [6] (Akamatsu, "Facial image matching device", Japanese Patent Laid-Open No. 5-20442) and Reference [7] (Akamatsu et al., "Robust front face identification method using grayscale image matching -Application of KL expansion of Fourier spectrum- In "Science Theory Vol.J76-D-II, No.7, 1993), the intensity distribution of the spatial frequency component of the reflected luminance image, that is, the Fourier spectrum is used as the feature quantity.
[0040]
The greatest merit of using the Fourier spectrum as a feature quantity is that it is not affected by the positional deviation fluctuation of the object image in the image pattern due to the invariance of the geometric transformation. Furthermore, it is shown in the literature [7] that the Fourier spectrum has the effect of mitigating the influence of illumination fluctuations. This is because the illumination fluctuation affects only the low frequency region of the Fourier spectrum, and the influence of the illumination fluctuation can be suppressed to a local fluctuation called intensity fluctuation in the low frequency region. This fluctuation in the low frequency region is a local fluctuation corresponding to (b) in differentiation. Therefore, in the case of the Fourier spectrum, there is a merit of using the mutual subspace method as a pattern recognition method as in the case of differentiation.
[0041]
As described above, both the differential image and the Fourier spectrum alone have the effect of localizing / miniaturizing the illumination fluctuation component from the reflected luminance image. Here, it is conceivable to further reduce the illumination variation component that remains after the initial illumination variation stabilization processing by these single units by another method. There is no example that clearly suggests the means for making the remaining illumination fluctuation clear, and a combination of the means for making the initial light fluctuation constant and the means for making the remaining light fluctuation permanent is realized. This is the second problem in the prior art.
[0042]
Furthermore, the low frequency component of the high reflectance part that strongly influences the illumination, that is, the shadow change that should be reduced as the illumination fluctuation component, includes a reflection luminance pattern due to the shape characteristic of the object. The high-pass filtering process for the reflected luminance image is an operation for dropping such feature information, and only the reflectance feature of the object is used for recognition. I hope to achieve high recognition performance by reducing the effects of lighting fluctuations while using both. This is the third problem in the prior art.
[0043]
Further, general differentiation processing requires at least a weighted sum calculation of pixel values, that is, a product-sum operation, and Fourier expansion requires much more calculation. Therefore, the calculation cost may increase unacceptably as the number of pixels in the image increases.
[0044]
In particular, the product-sum operation in differentiation is a basic method of realizing high-pass filtering, so speeding up this maximizes the processing performance of a pattern recognition system that combines high-pass filtering and the mutual subspace method. It is also important to improve to the limit. The realization of faster high-pass filtering is a fourth problem in the prior art.
[0045]
[Problems to be solved by the invention]
The present invention has been made in view of the above problems, and the first object of the present invention is to reduce the weakness of the mutual subspace method to illumination fluctuations by means of localizing / miniaturizing the illumination fluctuations. It is to provide a pattern recognition apparatus and method that are more robust than any other conventional methods for fusing the advantages of the two by making up for them, and for overall fluctuations such as fluctuations in the position and orientation of the object, deformation of the object, and illumination fluctuations.
[0046]
In addition, the second object of the present invention is to provide another means for further reducing the illumination fluctuation component still remaining after the means for localizing / minimizing the illumination fluctuation, so that the unlearned minute fluctuation appearing in the input is provided. Therefore, it is intended to realize illumination resistance variability that is more robust to the pattern recognition apparatus and method as the first object.
[0047]
A third object of the present invention is to provide a pattern recognition means that localizes / miniaturizes the influence of illumination variation while effectively using the object texture feature by the reflectance pattern and the object shape feature by the shadow pattern. By providing it, the object features appearing in the input can be used to the maximum extent, thereby improving the recognition capability of the pattern recognition apparatus and method as the first object.
[0048]
The fourth object of the present invention is to provide a means for localizing / minimizing illumination fluctuations by an extremely high speed and simple process, and thus the pattern recognition device as the first object and the processing capability of the method. It is to improve.
[0049]
[Means for Solving the Problems]
The pattern recognition apparatus according to the present invention includes means for inputting a pattern, means for extracting a feature vector for recognition from the input pattern, means for generating an input subspace from the feature vector, and the input subspace. In a pattern recognition apparatus comprising means for calculating similarity between dictionary subspaces and means for determining an object category based on the similarity, the feature vector localizes / miniaturizes illumination variation 1 One of the one or more processes which are extracted through a plurality of processes and localize / miniaturize the illumination variation is a process of expanding the range of the low luminance region and relatively compressing the range of the high luminance region. It is characterized by that.
[0050]
The pattern recognition apparatus according to the present invention includes means for inputting a pattern, means for extracting a feature vector for recognition from the input pattern, means for generating an input subspace from the feature vector, and the input subspace. In a pattern recognition apparatus comprising means for calculating a similarity between dictionary subspaces and means for determining an object category based on the similarity, a part or all of the feature vectors localize illumination variations / Extracted through one or more processes to be miniaturized, and one of the one or more processes to localize / miniaturize the illumination variation expands the range of the low-brightness region and increases the range of the relatively high-brightness region. The input subspace is generated through a process of projecting a subspace obtained by principal component analysis of the feature vector onto a constrained subspace that suppresses illumination variation, About subspace, characterized in that it is produced by the feature vector collected under the same lighting conditions as a linear combination of the difference subspace between categories of the signal space defined for each category.
[0051]
The pattern recognition apparatus according to the present invention includes means for inputting a pattern, means for extracting a first feature vector for recognition from the input pattern, and extracting a second feature vector for recognition from the input pattern. Means for generating an input subspace from the first and second feature vectors, means for calculating a similarity between the input subspace and the dictionary subspace, and an object based on the similarity In a pattern recognition apparatus comprising a means for determining a category, the first feature vector is extracted through one or more processes for localizing / minimizing illumination variation, and localizing / miniaturizing the illumination variation. One of the one or more processes is a process of expanding the range of the low luminance region and relatively compressing the range of the high luminance region.
[0052]
According to the present invention, in the pattern recognition device for recognizing the reflected luminance pattern obtained by photographing a three-dimensional object, the influence of illumination variation appearing in the pattern is often included in the low frequency component of the spatial frequency and the high luminance pixel region. As a process to use and localize / miniaturize illumination fluctuations, for example, high-pass filtering for patterns, low-intensity range expansion processing for patterns, and Fourier expansion processing for patterns can be performed alone or in combination to reduce the effects of illumination. Generate and recognize localized / miniaturized feature vectors. As a result, it is possible to realize a pattern recognition device that compensates for the weakness of the mutual subspace method with respect to illumination fluctuations, restores the original fluctuation resistance, and is robust against fluctuations in general.
[0053]
According to the present invention, in a pattern recognition apparatus for recognizing a reflected luminance pattern obtained by photographing a three-dimensional object, a feature vector is generated by performing processing for localizing / minimizing illumination fluctuations, and further, input by the feature vector By projecting the subspace onto a predetermined constrained subspace, a new input subspace that does not expand in the direction of illumination variation is generated. As a result, it is possible to realize a pattern recognition device that compensates for the weakness of the mutual subspace method with respect to illumination fluctuations and restores the original fluctuation resistance, and is more robust against fluctuations in general.
[0054]
According to the present invention, in the pattern recognition device for recognizing the reflected luminance pattern obtained by photographing a three-dimensional object, the first feature vector is generated by performing the processing for localizing / minimizing the illumination variation, and separately from this. An input subspace in which the second feature vector that is not subjected to the localization / miniaturization processing of the illumination variation is generated and the effect of illumination is reduced by appropriately combining both feature vectors, and many object-specific features are stored. Is generated. As a result, it is possible to realize a pattern recognition apparatus that compensates for the weakness of the mutual subspace method against illumination fluctuations and restores the original fluctuation resistance, and is robust against the fluctuations and improved in recognition performance.
[0055]
According to the present invention, in the pattern recognition device for recognizing a reflected luminance pattern obtained by photographing a three-dimensional object, the high-pass filtering process is realized by a differentiation process, particularly an embossing process with extremely low calculation cost. As a result, there is almost no increase in calculation cost due to the addition of this processing, and the weakness against the illumination fluctuation of the mutual subspace method is compensated to restore the original fluctuation resistance. A pattern recognition device can be realized.
[0056]
According to the present invention, by performing luminance correction on the pixel luminance value of the pattern prior to the process of localizing / miniaturizing the illumination variation, a feature vector that is more stable against the illumination variation is generated, and the mutual subspace is generated. It is possible to recover the original fluctuation resistance by compensating for the weakness of the law against illumination fluctuations, and to realize a pattern recognition device that is robust and stable with respect to all fluctuations.
[0057]
The invention relating to each of the above devices is also established as an invention relating to the method, and the invention relating to the method is also established as an invention relating to the device.
[0058]
The above-described invention can also be realized as a machine-readable medium recording a program for causing a computer to execute a corresponding procedure or means.
[0059]
DETAILED DESCRIPTION OF THE INVENTION
Embodiments of a pattern recognition apparatus and method according to the present invention will be described with reference to the drawings.
[0060]
(Description of the first embodiment)
A first embodiment of a pattern recognition apparatus and method according to the present invention will be described below. FIG. 1 shows a block configuration of the apparatus of this embodiment. The apparatus according to the present embodiment includes a pattern input unit 1, a fluctuation constant vector generation unit 2, an input subspace generation unit 3, an intersubspace similarity calculation unit 4, a recognition category determination unit 5, and an output unit 6. FIG. 2 shows a processing configuration in the apparatus of this embodiment. The processing in the apparatus of the present embodiment includes pattern input processing S1, variation constant vector generation processing S2, input subspace generation processing S3, subspace similarity calculation processing S4, recognition category determination processing S5, and output processing S6.
[0061]
(First Embodiment: Description of Pattern Input Unit 1)
The pattern input unit 1 inputs k reflection luminance patterns {p: p1,..., Pk} of a recognition object photographed using an imaging means such as a television camera (step S1). At this time, the pattern input unit 1 can input a plurality of patterns having different shooting times at a single imaging position, or can input a plurality of patterns having the same or different shooting times at a plurality of imaging positions. Is possible. The pattern input unit 1 may use a single imaging unit or a plurality of imaging units.
[0062]
(First Embodiment: Explanation of Fluctuation Constant Vector Generator 2)
The effect of illumination is to generate a shadow derived from the object shape on the reflection luminance pattern in the high reflectance portion. Since the variation in illumination invites a large-scale variation over the entire shadow pattern, it is necessary to suppress this large-scale variation to a local and minute variation so that the stability of the mutual subspace method is exhibited. The variation constant vector generation unit 2 is a functional block that localizes / miniaturizes such variation, and the influence of illumination variation from each reflected luminance pattern {p: p1, ..., Pk} by the pattern input unit 1. A feature vector {v: v1, ..., vk} obtained by localizing / minimizing is generated and output (step S2). For this purpose, the variation-constant vector generation unit 2 in the apparatus of the present embodiment uses at least the following three alone or in combination as a method for generating a feature vector in which the influence of illumination variation is reduced from the reflected luminance pattern. Can do.
[0063]
(A) High pass filtering
(B) Expanding the range of the low brightness area
(C) Fourier transform
Hereinafter, the methods (A) to (C) will be described.
[0064]
(First embodiment: description of the fluctuation constant vector generation unit 2: high-pass filtering)
As a first method for localizing / miniaturizing the illumination variation, high-pass filtering (A) for the reflected luminance pattern {p: p1,..., Pk} can be mentioned. By using the fact that the shadow variation that appears in the high reflectance part due to illumination variation is a low frequency component in terms of spatial frequency, high-pass filtering (or a reflection luminance pattern {p: p1, ..., Pk}) (or By generating an image pattern that has been subjected to high-frequency emphasis filtering as a feature vector {v: v1, ..., vk}, the shading variation is converted into inconsistent weak edges (local And miniaturization).
[0065]
Although several candidates are conceivable as a method for realizing high-pass filtering, the simplest method is to apply a differential process to the reflected luminance pattern {p: p1, ..., Pk}. In general, the differential processing is performed by scanning the differential operator Op of (2m + 1) × (2n + 1) pixels on the image pattern of w × h pixels, and the image pattern in which the weight coefficient of the pixel of operator Op overlaps with the operator pixel. A value (weighting coefficient × luminance value) multiplied by the luminance value of the upper pixel is obtained, and a value (sum of products of the weighting coefficient and the luminance value) obtained by adding such calculation values of all the operator pixels corresponds to the operator center. The differential value on the image pattern. Since the area covered by the operator is in the vicinity of the periphery of the operator center pixel on the image pattern, the operator is considered to be farther from the center pixel as the operator size is larger, that is, the operator detects a low spatial frequency component having a long wavelength. Conversely, in order to detect a high spatial frequency component, the operator should have a small size. Detailed information on image processing by a differential operator is described in the literature [8] (edited by Hiroshi Tanaka, “Image Processing Applied Technology”, Industrial Research Committee, 1989). The effective range of the differentiation process by the operator is the range obtained by removing the m pixels at both the left and right ends of the image pattern and the n pixels at the upper and lower ends, and the pattern after differentiation is (w-2m) x (h-2n) of this effective range. The pixel pattern, that is, a (w-2m) × (h-2n) -dimensional feature vector. This is because an operator whose center pixel is outside the effective range protrudes an image pattern, and thus a differential value cannot be obtained accurately.
[0066]
In addition to the weighted sum calculation by applying the differential operator Op, the differential processing can be realized by, for example, embossing processing known as a relief-like image expression technique. In embossing, the same image pattern is shifted and superimposed on the original image pattern, and the value obtained by subtracting the other from one is used as the value of the pixel. The embossing process can be regarded as a one-way differential process using an extremely small differential operator, but it can be performed by subtracting only the pixel values, so it is executed much faster than the weighted sum calculation by a general differential operator. it can. In particular, the embossing process that is performed by shifting one pixel functions as a high-pass filter that extracts only the highest spatial frequency component in the reflection luminance pattern, so that low-frequency components that are affected by illumination fluctuations can be removed with low calculation cost. Convenient to do.
[0067]
In the case of embossing, the range where the image patterns overlap is the effective range, so the feature vector is also a vector of the number of pixels in the overlapping range. Embossing by shifting by one pixel is more advantageous because the effective range has only one pixel width at most, and the effective range can be kept wider than when a large differential operator is applied.
[0068]
The direction in which the image pattern is shifted in the embossing process is most effective if it is perpendicular to the direction of the edge representing the feature of the recognition target appearing in the image pattern. By doing in this way, the illumination variation constant pattern becomes an image in which the characteristics of the recognition target remain sufficiently and the influence of illumination variation is eliminated as much as possible. For example, since human faces have many horizontal edges such as eyebrows, eyes, and mouths, embossing by shifting the image pattern in the vertical direction (the vertical axis direction of the face) is considered effective. The same directionality problem applies to ordinary differential operators, and it is important to apply sensitive differential operators only in the important edge directions representing the characteristics of the pattern.
[0069]
The image pattern generated and output by high-pass filtering has a strong edge (a) corresponding to the reflectance pattern of the low reflectance portion and a weak edge (b) corresponding to the shadow of the high reflectance portion. Yes. Of these, the former (a) is a component to be recognized as a solid image feature representing an object feature, while the latter (b) is a noise component whose position and intensity change due to illumination variation. Although the objective of localizing / minimizing the illumination variation over the entire pattern to the noise component as shown in (b) can be achieved only by the high-pass filtering, this residual is followed by the high-pass filtering (constant illumination variation). By adding a process for removing the noise component (b) (constant residual illumination fluctuation), the illumination fluctuation is further reduced, and the fluctuation stability of the mutual subspace method is further effective. Therefore, weak edge removal processing is executed to remove the noise component (b) as the residual illumination fluctuation constant processing.
[0070]
Specifically, a pixel value whose absolute value is equal to or less than a predetermined threshold value is replaced with 0. The image pattern generated as a result is a feature vector that leaves only the edge of (a).
[0071]
(First embodiment: description of the variation constant vector generation unit 2: expansion of the range of the low luminance region)
As a second method for localizing / miniaturizing the illumination variation, there is a range expansion process (B) in the low luminance region for the reflected luminance pattern {p: p1,..., Pk}. Utilizing the fact that fluctuations in lighting do not affect the low-reflectance part so much, the brightness range of the low-brightness area is expanded in order to emphasize the image information of the low-reflectance part that always has low reflection brightness in the reflection brightness pattern. Reducing the luminance range of the relatively high luminance area. That is, the gradation of the dark part in the image pattern is widened and the gradation of the bright part is narrowed. As a result, the information of the high reflectance portion that is easily affected by illumination is suppressed (miniaturized) to a narrow gradation range, and the information of the low reflectance portion that strongly retains the characteristics of the object is re-expressed in a wide gradation range. As a specific calculation method, a brightness correction curve that defines an output brightness value with respect to input brightness and a brightness correction table that stores an output brightness value with respect to input brightness are prepared, and a corrected brightness value with respect to input brightness is obtained.
[0072]
Also, as an extreme example of the process of enlarging the luminance range of the low luminance region, a method of binarizing the reflected luminance pattern is also possible. The pixel value of the reflection luminance pattern is compared with a predetermined threshold value, and if the luminance is higher than the threshold value, a predetermined value is given to the pixel, and if the luminance is lower than the threshold value, another predetermined value is given to the pixel.
[0073]
In either method, since the effective range of the pattern after processing does not change, the dimension of the feature vector matches the number of pixels w × h of the input pattern.
[0074]
(First embodiment: description of the fluctuation constant vector generation unit 2: Fourier expansion)
As a third method for localizing / miniaturizing the illumination variation, Fourier expansion (C) for the reflection luminance pattern {p: p1,..., Pk} disclosed in the literature [7] can be mentioned. Since the shading fluctuation accompanying the lighting fluctuation is a low frequency component in terms of the spatial frequency, the influence of the lighting fluctuation is the spatial frequency obtained by Fourier expansion of the reflection luminance pattern {p: p1, ..., Pk}. It often appears in the low frequency region of another intensity (Fourier spectrum). Each Fourier spectrum {f: f1, ..., fn} generated for n spatial frequencies of k patterns is regarded as one of N-dimensional feature vectors {v: v1, ..., vk}, respectively. The effect of illumination variation appears only in a small number of scalars corresponding to the low frequency region within it (localization).
[0075]
Note that the feature vector generated by Fourier expansion is a Fourier spectrum composed of N spectral components. The Fourier spectrum expresses the image characteristics of an object with a pattern of spatial frequency components, and the influence of shadows due to illumination is still preserved in the low frequency components. The purpose of localizing the illumination variation over the entire pattern to the low-frequency component of the Fourier spectrum can be achieved by Fourier expansion alone, but following the Fourier expansion (constant illumination variation), this remaining low-frequency component (b) is removed. By adding this processing (residual illumination fluctuation constant), the illumination fluctuation is further reduced, and the fluctuation stability of the mutual subspace method is more effective. Therefore, as a residual illumination fluctuation constant processing, a feature vector is reconstructed with (Nn) spectral components obtained by deleting n spectral components corresponding to low frequency components from among N spectral components. As a result, a feature vector from which the influence of illumination variation is removed is obtained.
[0076]
(First Embodiment: Explanation of Fluctuation Constant Vector Generator 2: Luminance Correction / Normalization Processing)
In addition, prior to high-pass filtering (A), low-brightness range expansion processing (B), and Fourier expansion (C), the variation of the reflected luminance pattern {p: p1, ..., Pk} Correct and normalize the brightness range and brightness histogram. By doing so, the edge intensity of the differential image, the low luminance region corresponding to the low reflectance portion, and the intensity of the Fourier spectrum can be stably obtained regardless of the fluctuation of the illumination intensity. As the luminance correction / normalization processing, norm normalization with respect to the reflection luminance pattern (a common coefficient is applied to each pixel value so that the square sum of the pixel values becomes a predetermined value), canonicalization (the average of the pixel values is 0, a common offset is subtracted from each pixel value so that the variance of the pixel values is 1, and a common coefficient is applied), and histogram flattening (the luminance value of each pixel is made nonlinear so that the luminance histogram becomes flat) Correction) and the like.
[0077]
(First embodiment: description of the fluctuation constant vector generation unit 2: combination of feature amounts, etc.)
In addition, the method for lighting fluctuation constant is not limited to the example of said (A)-(C). For example, CCCI (Color Constant Color) disclosed in the literature [9] (Hashizume et al., “Object extraction from color images robust to illumination light fluctuations”, IEICE Technical Report PRMU97-125, 1997) is proposed as a constant lighting fluctuation. Indexing) descriptors can also be used. Specifically, the weighted sum of the logarithm of the luminance value of each pixel and the logarithm of the luminance value in the vicinity of the four adjacent pixels is used. This utilizes the property (assuming) that the luminance ratio between multiple points in the reflected luminance pattern does not change even if the illumination varies, and has the effect of minimizing the influence of illumination variation. In short, in the present invention, the mutual subspace method exhibits resistance to rare fluctuations and minute fluctuations due to the synergistic effect of the stability of the input subspace and the stability of the subspace similarity, and thus the recognition system has almost all fluctuations. As long as the illumination fluctuation that affects the entire area of the feature vector can be suppressed to local fluctuations and minute fluctuations so that high stability can be obtained, differences in the means for localizing and minimizing fluctuations Is not a problem.
[0078]
Further, the feature vector generated through the illumination variation constant (or the process with the residual illumination variation constant) is generated by using a plurality of the above-described (A) to (C) and other corresponding methods in parallel. It is also possible to make it. That is, it is possible to generate and use a new feature vector by appropriately combining each scalar quantity of feature vectors generated by a plurality of methods. For example, each feature quantity by high-pass filtering and Fourier expansion may be edited into one feature vector and used. In this way, it is possible to achieve stable recognition by localizing / miniaturizing the illumination variation included in the object using a feature expression that cannot be obtained by one method.
[0079]
(First Embodiment: Description of Input Subspace Generation Unit 3)
The input subspace generator 3 uses the eigenvalue {λ: λ1,... By the principal component analysis technique such as KL expansion from the N-dimensional feature vector {v: v1,. , λM} and N-dimensional eigenvectors {e: e1, ..., eM}, and Mi (Mi ≦ M) eigenvectors {e: e1, ..., eMi} in descending order of corresponding eigenvalues. An input subspace Li (dimension number Mi) is generated and output as a basis vector (step S3). This Mi is the number of bases on the input side used for similarity calculation in the subspace similarity calculation unit 4 in the next stage.
[0080]
(First embodiment: description of similarity calculation unit 4 between subspaces)
The inter-subspace similarity calculation unit 4 holds dictionary subspaces corresponding to x categories to be recognized, and x dictionary subspaces held by the input subspace generation unit 3 and the input subspace Li. The subspace similarity {s: s1, ..., sx} with {Ld: Ld1, ..., Ldx} is calculated, and the similarity and the category code {c : C1, ..., cx} is generated and output as recognition candidate information {c: c1, ..., cx, s: s1, ..., sx} (step S4). When the projection matrix for the input subspace Li is P and the projection matrix for the dictionary subspace Ld is Q, the subspace similarity s is the maximum eigenvalue of the matrix PQP or QPQ. At this time, Mi basis vectors of the input subspace Li are used to generate the projection matrix P 1, and Md basis vectors of the dictionary subspace Ld are used to generate the projection matrix Q 1. The definition of similarity between subspaces and the mathematical support of the calculation method are disclosed in the literature [3].
[0081]
(First Example: Explanation of Recognition Result Determination Unit 5)
The recognition result determination unit 5 compares all the similarities of the recognition candidate information {c: c1,..., Cx, s: s1,. The category code co in the dictionary subspace showing the above highest similarity is confirmed and output as a recognition result. If there is no dictionary subspace indicating a similarity equal to or greater than a predetermined threshold, it is determined that a recognizable category is not included in the input pattern, and a special code indicating no category is output (step S5). .
[0082]
(First embodiment: description of output unit 6)
The output unit 6 outputs at least the code co from the recognition result determination unit 5 to the outside of the apparatus (step S6).
[0083]
(Description of the second embodiment)
The first implementation of the embodiment in which the reflected luminance pattern in which the influence of illumination variation is localized / miniaturized through the variation constant vector generation unit 2 (variation constant vector generation processing S2) is recognized by the mutual subspace method. It was an example explanation. Next, a pattern recognition apparatus and method according to a second embodiment of the present invention will be described.
[0084]
FIG. 3 shows a block configuration of the apparatus of this embodiment. The apparatus according to the present embodiment includes a pattern input unit 11, a variation constant vector generation unit 12, an input subspace generation unit 13, a subspace similarity calculation unit 14, a recognition category determination unit 15, an output unit 16, and a subspace projection unit 17. Become. FIG. 4 shows a processing configuration in the apparatus of this embodiment. The processing in the apparatus of the present embodiment is pattern input processing S11, variation constant vector generation processing S12, input subspace generation processing S13, subspace similarity calculation processing S14, recognition category determination processing S15, output processing S16, subspace projection processing. Consists of S17. 11 to 16 (S11 to S16) in the figure are the same as 1 to 6 (S1 to S6) in the first embodiment. The difference between the present embodiment and the first embodiment is the addition of the subspace projection unit 17 (subspace projection processing S17).
[0085]
(Second embodiment: description of subspace projection unit 17)
The subspace Li generated by the input subspace generation unit 13 is a subspace with the dimension number Mi generated by the feature vector v in which the variation in illumination is localized / miniaturized by the variation constant vector generation unit 12. The effects of fluctuations are not completely eliminated. The subspace projection unit 17 holds a constrained subspace Lc having a dimension number Mc, and generates and outputs a new subspace Li ′ obtained by projecting the input subspace onto the constrained subspace Lc (step S17). At this time, the calculation for projecting the input subspace Li to the constrained subspace Lc is to obtain new Mi vectors obtained by projecting Mi base vectors of the input subspace Li to Lc, respectively, and orthonormalize this vector to obtain a partial This is achieved by using the base of the space Li ′. In addition, the constrained subspace Lc that constrains illumination fluctuations collects x categories of M-dimensional subspaces {Ls: Ls1,..., Lsx} under the same lighting conditions, and between any two different categories. y differential subspaces {LD: LD1, ..., LDy} are obtained and given as a linear combination thereof.
[0086]
The difference subspace LD is given with a difference vector (uv) of M sets of vectors u and v giving canonical angles of the two subspaces Lsi and Lsj as a basis. Since the differential subspace {LD: LD1,..., LDy} generated under the same illumination condition does not include illumination variation, the constrained subspace Lc, which is a linear combination thereof, does not expand in the illumination variation direction. On the other hand, the difference subspace {LD: LD1,..., LDy} is a difference between categories, so that the constrained subspace Lc that is a linear combination also has a spread in the direction of the difference between categories.
[0087]
Therefore, the input subspace Li projected onto the constrained subspace Lc does not include a component in the illumination variation direction, and becomes an input subspace Li ′ that stores the component in the difference direction between categories. In other words, the partial space after projection is a partial space in which variation in illumination is suppressed and category features are emphasized. Note that the important point here is that the collection of {Ls: Ls1,..., Lsx} for obtaining the constrained subspace Lc can be easily performed because the illumination conditions may be constant.
[0088]
The inter-subspace similarity calculation unit 14 holds a dictionary subspace {Ld ′: Ld1 ′,..., Ldx ′} previously projected onto the constrained subspace Lc, and an input portion by the subspace projection unit 17 Compute all subspace similarity {s: s1, ..., sx} between the space Li ′ and the x dictionary subspaces {Ld ′: Ld1 ′,.. Recognition candidate information {c: c1, ..., cx, s: s1, ... paired with the category code {c: c1, ..., cx} of the dictionary subspace indicating the similarity. , sx} is generated and output (step S14). The recognition method for calculating the similarity between subspaces after projecting the input subspace and the dictionary subspace to the constrained subspace as described above is called a constrained mutual subspace method.
[0089]
By using the illumination variation constant in the variation constant vector generation unit 12 (or processing with constant residual illumination variation constant) and the constrained mutual subspace method by the subspace projection unit 17, similarity that is not affected by illumination variation can be obtained. It can be used for calculation.
[0090]
(Explanation of the third embodiment)
The above is an embodiment using a feature vector subjected to illumination fluctuation constant. However, in the present invention, a feature vector generated through illumination fluctuation constant and a feature generated without going through illumination fluctuation constant. It is also possible to use a combination of vectors.
[0091]
In other words, by generating and using a new feature vector that appropriately combines each scalar quantity of each feature vector, the object feature information that can no longer be used due to the constant illumination variation is used from the feature vector that is not subject to the constant illumination variation. In addition, in the feature vector in which both are made possible, the scale of variation is localized / miniaturized as an effect of making the illumination variation constant as compared with the case where the illumination variation is not made constant. As a result, it is possible to perform stable recognition by reducing the illumination variation included in the object while using the feature expression of the object that cannot be obtained by making the illumination variation constant. The pattern recognition apparatus and method according to the third embodiment of the present invention is given such a configuration. The third embodiment will be described below.
[0092]
FIG. 5 shows a block configuration of the apparatus of this embodiment. The apparatus according to the present embodiment includes a pattern input unit 21, a variation constant vector generation unit 22, an input subspace generation unit 23, a subspace similarity calculation unit 24, a recognition category determination unit 25, an output unit 26, and a variation non-constant vector generation. Part 27. FIG. 6 shows a processing configuration in the apparatus of this embodiment. The processing in the apparatus of the present embodiment is pattern input processing S21, variation constant vector generation processing S22, input subspace generation processing S23, subspace similarity calculation processing S24, recognition category determination processing S25, output processing S26, variation non-constant. It consists of vector generation processing S27. 21 to 26 (S21 to S26) in the figure are the same as 1 to 6 (S1 to S6) in the first embodiment. The difference between the present embodiment and the first embodiment is the addition of a fluctuation non-constant vector generation unit 27 (fluctuation non-constant vector generation process S27).
[0093]
The variation constant vector generation unit 22 is a first N1-dimensional feature obtained by performing illumination variation constant (or processing including residual illumination variation constant) by, for example, high-pass filtering from the reflected luminance pattern by the pattern input unit 21. Generate a vector. At this time, the first feature vector is a vector in which illumination variation is localized / miniaturized, and most of the shadow information that expresses the shape feature of the object is lost, and only strong edge information that expresses the texture feature of the object is lost. Is maintained.
[0094]
The variation non-constant vector generation unit 27 directly generates an N2-dimensional second feature vector from the reflected luminance pattern by the pattern input unit 21. The second feature vector is the reflection luminance pattern itself, and the shadow information based on the object shape is maintained as it is, but this also includes the influence of illumination variation over the entire pattern.
[0095]
The input subspace generation unit 23 considers the N1 dimensional first feature vector and the N2 dimensional second feature vector as a (N1 + N2) dimensional feature vector and inputs a plurality of (N1 + N2) dimensional inputs. An input subspace is generated from the feature vector.
[0096]
In the present embodiment, the first feature vector is generated by making the illumination variation constant (or processing adding the residual illumination variation constant) only for the first predetermined region of the pattern, and the second feature vector. Is generated by other means targeting only the second predetermined region of the pattern, and it is possible to perform recognition in a state where other unnecessary information is removed while leaving more object features. For example, in face recognition, the first predetermined area may be an eyebrow, eye, nose, or mouth periphery with many strong edges, and the second predetermined area may be a forehead or cheek. By doing this, the illumination variation is made constant in an area where many texture features are stored, and more face feature information is saved without making the illumination variation constant in an area where many solid shape features are stored. However, it is possible to use a feature amount with reduced illumination variation.
[0097]
(Description of Modified Example)
It should be noted that the embodiments described above can be implemented by combining or modifying the configurations.
[0098]
For example, the first feature vector is generated by making the illumination variation constant (or processing with constant residual illumination variation constant), the second feature vector is generated by other means, and the feature vector obtained by merging the two is used. It is also possible to generate the input subspace and perform projection onto the constrained subspace to obtain the similarity. In this way, it is possible to perform recognition by removing illumination variations more strongly while leaving more object features.
[0099]
In addition, the first predetermined region is a region having many strong edges such as the eyebrows, eyes, nose, and mouth around the face, and the second predetermined region is, for example, the forehead or cheek A first feature vector is generated as an area having a strong shadow by performing illumination fluctuation constant (or processing including residual illumination fluctuation constant) for only the first predetermined area, and the second feature vector is generated as the second feature vector. It is also possible to generate the input subspace from the feature vector that combines only the predetermined area of the target and generate the input subspace from the feature vector and then project it to the constrained subspace to obtain the similarity Is possible. In this way, it is possible to perform recognition in a state where illumination variation and other useless information are removed while leaving more object features.
[0100]
(Description of embodiment by medium)
By the way, as illustrated in FIG. 7, information (for example, a program) for realizing the image pattern recognition apparatus and method according to the present invention is recorded on a recording medium 31, and the recorded information is transmitted via the recording medium 31. The present invention can also be applied to the apparatus 32 and the apparatus 33 including the imaging unit 37, or to the apparatus 34 including the apparatus 33 and the imaging unit 38 via the communication lines 35 and 36. That is, the present invention is not limited to the above-described embodiment, and can be implemented with various modifications within the technical scope thereof.
[0101]
【The invention's effect】
According to the present invention, it is possible to recognize an image pattern that is not easily affected by the illumination variation even under an illumination variation environment that is affected by outside light such as outdoors or indoors on a window.
[0102]
Specifically, according to the present invention, in a pattern recognition device for recognizing a reflected luminance pattern obtained by photographing a three-dimensional object, the influence of illumination variation that appears in the pattern is applied to a low frequency component of a spatial frequency or a high luminance pixel region. A feature vector in which the influence of illumination is localized / miniaturized is generated and recognized by the illumination fluctuation constant processing utilizing the fact that it is included in large numbers. As a result, it is possible to realize a pattern recognition device that compensates for the weakness of the mutual subspace method with respect to illumination fluctuations, restores the original fluctuation resistance, and is robust against fluctuations in general.
[0103]
Further, according to the present invention, in the pattern recognition device for recognizing the reflected luminance pattern obtained by photographing a three-dimensional object, the illumination fluctuation variation is performed by performing the process of removing the fluctuation component remaining after the illumination fluctuation stabilization process. It is possible to generate a feature vector that further reduces the influence of the pattern recognition, and it is possible to realize a pattern recognition device that is more robust against fluctuations in general.
[0104]
Further, according to the present invention, in a pattern recognition device for recognizing a reflected luminance pattern obtained by photographing a three-dimensional object, a feature vector is generated by performing illumination variation stabilization processing, and the remaining illumination variation is further converted into a partial space. By performing the constrained mutual subspace method that becomes constant by projection, it is possible to realize a pattern recognition device that is even more robust against fluctuations in general.
[0105]
According to the present invention, in the pattern recognition device for recognizing the reflected luminance pattern obtained by photographing a three-dimensional object, the illumination variation constant processing is performed to generate the first feature vector. By generating and combining second feature vectors that are not subject to regularization, it is possible to generate feature vectors that preserve many object-specific features while localizing / minimizing illumination variations, and are robust and aware of all variations A pattern recognition device with improved performance can be realized.
[0106]
Further, according to the present invention, in the pattern recognition device for recognizing the reflected luminance pattern obtained by photographing the three-dimensional object, the illumination variation constant processing is realized by embossing processing with extremely low calculation cost. It is possible to realize a pattern recognition device that is robust and robust against all fluctuations with almost no increase in calculation cost due to the addition of the digitizing process.
[0107]
In addition, according to the present invention, by correcting the pixel luminance value of the pattern prior to the illumination variation stabilization process, it is possible to generate a feature vector that is more stable with respect to illumination variation, and is robust against all variations. A stable pattern recognition device can be realized.
[Brief description of the drawings]
FIG. 1 is a diagram showing a block configuration of an apparatus according to a first embodiment of the present invention.
FIG. 2 is a diagram showing a processing configuration of the first embodiment apparatus according to the present invention;
FIG. 3 is a diagram showing a block configuration of a second embodiment apparatus according to the present invention;
FIG. 4 is a diagram showing a processing configuration of a second embodiment apparatus according to the present invention.
FIG. 5 is a diagram showing a block configuration of a third embodiment apparatus according to the present invention.
FIG. 6 is a diagram showing a processing configuration of a third embodiment apparatus according to the present invention.
FIG. 7 is a diagram for explaining an embodiment using a recording medium or the like.
FIG. 8 is a diagram for explaining the effect of the mutual subspace method.
FIG. 9 is a diagram for explaining the influence of illumination variation.
FIG. 10 is a diagram for explaining the effect of localization / miniaturization of illumination variation.
[Explanation of symbols]
1: Pattern input unit,
2: Fluctuation constant vector generator,
3: Input subspace generator,
4: Subspace similarity calculation unit,
5: Recognition category determination section,
6: Output section

Claims (6)

Means for inputting a pattern; means for extracting a feature vector for recognition from the input pattern; means for generating an input subspace from the feature vector; and similarity between the input subspace and the dictionary subspace. In a pattern recognition device comprising means for calculating and means for determining a category of an object based on the similarity,
The feature vector is extracted through one or more processes for localizing / miniaturizing illumination fluctuations, and one of the one or more processes for localizing / miniaturizing the illumination fluctuations expands the range of the low luminance region. A pattern recognition apparatus, which is a process of compressing a range of a relatively high luminance region. Means for inputting a pattern; means for extracting a feature vector for recognition from the input pattern; means for generating an input subspace from the feature vector; and similarity between the input subspace and the dictionary subspace. In a pattern recognition device comprising means for calculating and means for determining a category of an object based on the similarity,
A part or all of the feature vector is extracted through one or more processes for localizing / miniaturizing illumination fluctuations, and one of the one or more processes for localizing / miniaturizing the illumination fluctuations is a low luminance region. Is a process that expands the range and compresses the range of the relatively high brightness area,
The input subspace is generated through a process of projecting a subspace obtained by principal component analysis of the feature vector onto a constrained subspace that suppresses illumination variation, and the constrained subspace is collected under the same lighting condition. A pattern recognition apparatus, wherein the pattern recognition device is generated as a linear combination of difference subspaces between categories of subspaces extended for each category by the feature vector. Means for inputting a pattern; means for extracting a first feature vector for recognition from the input pattern; means for extracting a second feature vector for recognition from the input pattern; and A pattern comprising: means for generating an input subspace from two feature vectors; means for calculating a similarity between the input subspace and the dictionary subspace; and means for determining an object category based on the similarity. In the recognition device,
The first feature vector is extracted through one or more processes for localizing / miniaturizing illumination fluctuations, and one of the one or more processes for localizing / miniaturizing the illumination fluctuations is a range of a low luminance region. The pattern recognition apparatus is a process for enlarging the image and compressing the range of the relatively high luminance region. A process of inputting a pattern, a process of extracting a feature vector for recognition from the input pattern, a process of generating an input subspace from the feature vector, and a similarity between the input subspace and the dictionary subspace In a pattern recognition method comprising a process of calculating and a process of determining a category of an object based on the similarity,
The feature vector is extracted through one or more processes for localizing / miniaturizing illumination fluctuations, and one of the one or more processes for localizing / miniaturizing the illumination fluctuations expands the range of the low luminance region. A pattern recognition method, which is a process of compressing a range of a relatively high luminance region. A process of inputting a pattern, a process of extracting a feature vector for recognition from the input pattern, a process of generating an input subspace from the feature vector, and a similarity between the input subspace and the dictionary subspace In a pattern recognition method comprising a process of calculating and a process of determining a category of an object based on the similarity,
The feature vector is extracted through one or more processes for localizing / miniaturizing illumination fluctuations, and one of the one or more processes for localizing / miniaturizing the illumination fluctuations expands the range of the low luminance region. Is a process of compressing the range of the relatively high brightness area,
The input subspace is generated through a process of projecting a subspace obtained by principal component analysis of the feature vector onto a constrained subspace that suppresses illumination variation, and the constrained subspace is collected under the same lighting condition. A pattern recognition method, wherein the pattern recognition method is generated as a linear combination of difference subspaces between categories of subspaces extended for each category by the feature vector. A process of inputting a pattern, a process of extracting a first feature vector for recognition from the input pattern, a process of extracting a second feature vector for recognition from the input pattern, the first and first A pattern comprising: a process of generating an input subspace from two feature vectors; a process of calculating a similarity between the input subspace and the dictionary subspace; and a process of determining an object category based on the similarity In the recognition method,
The first feature vector is extracted through one or more processes for localizing / miniaturizing illumination fluctuations, and one of the one or more processes for localizing / miniaturizing the illumination fluctuations is a range of a low luminance region. The pattern recognition method is a process of enlarging the image and compressing the range of the relatively high luminance region.

Images