JPH0786927B2 - Vector identification device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION [Industrial field of application] The present invention relates to images such as characters and other figures, various types of sound,
Classification such as class classification, recognition, etc. is performed on a vector consisting of single or multiple vector components corresponding to the characteristics of predetermined information such as technical information to be searched, and an image,
The present invention relates to a vector identification device adapted to perform classification, recognition and the like on voice, technical information to be searched and the like.

[Outline of Invention]

The value of each component of the input vector is coded to cause the linear display to emit light, and the reference vector represented by the transmissive / non-transmissive code string is orthogonal to the array direction of the strands of the linear display. By interposing a plurality of reference masks arranged in the direction with the two-dimensional photosensor array, or by using the two-dimensional photosensor array itself as a reference vector to form an array structure of light receiving / non-light receiving parts,
This is a vector identification device in which one input vector and a large number of reference vectors are collated by optical parallel processing, and vector identification can be performed without performing digital arithmetic processing.

[Conventional technology]

2. Description of the Related Art Conventionally, image recognition devices that recognize images such as characters and other figures are mainly configured to perform digital processing using electronic technology. Then, which of the reference data among the many reference data stored in the database is most matched with the data corresponding to the characteristics of the image to be recognized is compared by a correlation calculation.

However, in the case of character recognition in which the image to be recognized is, for example, a character, even when recognizing printed characters with a well-arranged font, Kanji has many strokes and is complicated, and there are thousands of character types.
Moreover, there are various types of fonts such as Cosic and Mincho even if they are called printed characters. In the case of handwritten characters, there are various variations depending on the person who wrote the character.
Therefore, in any case, the image recognition operation in the image recognition apparatus becomes very complicated, and it is very difficult to improve the accuracy while keeping the high speed operation.

On the other hand, among a large number of characters, there are many characters that are similar to each other as a whole and characters that are partly the same or similar to each other, and can be classified into several classes.

Therefore, in the image recognition apparatus, conventionally, the features of the image to be recognized are first extracted by projection or the like, and then the features such as the projection features are compared with the projection features of the reference pattern for class classification prepared in advance. By doing so, class classification is performed. Then, with respect to the images classified in this way, the image is specified, that is, the image is recognized, by a correlation calculation process in which correlation is made with a large number of images belonging to the class.
In the above-mentioned correlation calculation step, by comparing the projection features of the image to be recognized with the projection features of a large number of image recognition reference patterns prepared in advance,
A large number of standard patterns are ranked according to the degree to which the images to be recognized match, and the images are specified according to this ranking.

In the following, to further specifically supplement the description of the conventional example of the above-described image recognition apparatus configured to be digitally processed by mainly using the electronic technology, first, it is described by printing on an original or the like. An image pattern to be recognized is formed by an optical lens on the light receiving surface of an image sensor including a CCD, a MOS sensor and the like. Then, a multivalued digital signal as image information is output from the image sensor, and the multivalued digital signal is binarized with an appropriate threshold value (if there are a plurality of threshold values, the above-mentioned It is stored in the memory after performing multi-value quantization different from the case). Further, the binarized image information is stored in the memory or another memory after being subjected to preprocessing for shaping the image as necessary. By the pre-processing, noise removal processing and normalization processing of position, size, inclination, line width, etc. are performed.

Next, the projection processing unit extracts the projection feature from the image information stored in the memory in this way.

That is, when the image is projected onto a certain direction axis (for example, the X axis), the memory in which the image information is stored has a certain relationship with the direction axis and the certain direction (for example, the Y axis direction). To read the image information from the memory in time series or parallel time series, transfer the read image information to the projection processing unit, and then transfer the transferred image information to the projection processing unit. Accumulate sequentially. Then, the integrated value sequentially obtained by such integration is stored in a predetermined position corresponding to the certain direction axis in the above-mentioned memory or further memory. Further, based on the integrated amount stored in this way, the waveform of the intensity distribution from which the projective feature for the certain direction axis is extracted is obtained.

Since such a projective feature extraction operation of an image is usually performed for a plurality of directional axes, a plurality of intensity distribution waveforms can be obtained for the same image information. Then, the plurality of projection features of the image indicated by these intensity distribution waveforms are compared with the projection features of the reference pattern prepared in advance, so that the image classification or recognition is performed.

When digitally performing the above projective feature processing,
The integrated value, which is the input data, is regarded as one vector component, and one or a plurality of intensity distributions are treated as one vector having a large number of vector components. Therefore, for one intensity distribution,
If the integrated value is stored in n, this intensity distribution will be an n-dimensional vector.

In the image recognition apparatus as described above, in order to increase the image recognition rate, it is necessary to perform projection processing on a plurality of directional axes with respect to the same image information to extract a large number of projection features. Therefore, the image to be recognized is treated as an aggregate of multidimensional vectors in which many multidimensional vectors are aggregated.

On the other hand, the reference patterns for class classification and image recognition described above are also vectorized in the same manner as in the above case. Then, with a dedicated vector operation machine that has been put to practical use in parallel pipeline type computers, the correlation operation between the vector corresponding to the feature of the image to be recognized and the vector of many reference patterns is performed for each vector component. Be seen.

In the above correlation calculation, for example, the distance between the two vectors, the correlation coefficient, the degree of similarity, and the like are calculated as factors indicating the degree of correlation. Then, when determining the degree of correlation, it is necessary to search for an optimum correlation in consideration of variations in characteristics such as projection characteristics. Therefore, in reality, the correlation calculation between many vectors is performed many times by time-series digital processing.

[Problems to be solved by the invention]

However, in the conventional image recognition apparatus as described above, it is necessary to vectorize a large number of reference patterns for class classification or image recognition in advance. In addition, the correlation calculation between the vector corresponding to the feature of the image to be recognized and a large number of these vectorized reference patterns is performed by time-series digital processing in the vector calculation dedicated machine as described above. Must be done many times each time.

Therefore, a large amount of processing time is required for the vectorization and the correlation calculation, and a dedicated processor using a dedicated LSI needs to be incorporated in the dedicated vector calculation machine, resulting in a complicated device configuration. Therefore, there is a problem that the cost becomes high.

The present invention has been made in view of the above problems, and is simple by introducing optical parallel processing for identifying a vector composed of vector components corresponding to characteristics of predetermined information such as characters. With such a configuration, it is possible to identify a vector at high speed.

[Means for solving problems]

The vector identification device of the present invention includes a linear display 1 that encodes the values of the respective components of an input vector and emits light as a sequence of a plurality of line elements, and a translucent portion 3a and a non-translucent portion 3b corresponding to each linear element. A code string consisting of and extends in the array direction of the line elements to represent one reference vector, and the code strings of a plurality of different reference vectors are arrayed in a direction orthogonal to the array direction of the line elements. It comprises a reference mask 3 and a two-dimensional photosensor array 4 that faces the reference mask 3 and receives light from the linear display that passes through all of the light transmitting portions 3a.

Based on the output signal Y _i of the photosensor array 4 and the code value y _i of the input vector, the input vector and one of the plurality of different reference vectors X _i are identified.

The two-dimensional photosensor array itself may include the function of the reference mask. In this case, the sensor array is composed of a code string composed of a light receiving part corresponding to the reference vector and a substantially non-light receiving part.

[Action]

The input vector and a large number of reference vectors are collated instantaneously by optical parallel processing. No digital arithmetic processing is required for matching, and if the detection signal of the light passing through the reference mask is electrically discriminated whether it matches the code value of the input vector, the vector with the matched code value is extracted. The unknown input vector can be identified.

〔Example〕

Next, an embodiment in which the vector identification device of the present invention is applied to an image recognition device will be described with reference to FIGS.

First, a vector consisting of vector components corresponding to features such as projective features extracted from a character image to be recognized as preprocessing is obtained in the same manner as in the case of the conventional image recognition apparatus as described above.

FIGS. 2A and 2B respectively show intensity distribution waveforms by the X-axis direction projection and the Y-axis direction projection extracted from the character “mountain”, as an example. In this example, the vector corresponding to the projection feature of the character “mountain” includes five vector components C ₁ to 5 corresponding to each projection data of the X-axis direction projection feature,
It is composed of five vector components C ₆ to ₁₀ corresponding to the respective projection data of the Y-axis direction projection feature.

The projective feature may be an oblique direction, a circumferential direction, a radial direction, etc. in addition to the X and Y directions. In reality, there are about 16 vector components for each projection type. The distribution intensity value which is each vector component takes a value of 0 to 100, for example.

In the present invention, the intensity values of the individual vector components of the projection feature vector are encoded, and then optically identified in parallel with various reference codes for vector identification.

FIG. 1 shows a cross section of the main part of the vector identification device. The linear display 1 is formed by arranging a large number of linear light emitting sources L ₁ , L ₂ ... In parallel on the substrate 1a in a direction orthogonal to the plane of the drawing. The distribution intensity value which is each component of the projection feature vector is encoded and displayed on the linear display 1.

FIG. 3 shows a state in which one component of a vector is coded and displayed on the linear display 1. In this case, one component having a value of 0 to 100 is represented by 5 bits (5, 10, 15, ...) By 5 bits (5 line elements). The linear light source L _{1 at the} left end displays 0 or 5 at the first digit of the intensity value by turning off or stippling, respectively. The linear light emitting sources L _{2 to} L ₅ adjacent to L ₁ display 10 digits of the distribution intensity value in a 4-bit binary coded decimal system. In the example of FIG. 3, the diagonal lines indicate lighting, and each example has an intensity value of 65 (left),
Intensity values of 100 (right) are shown.

The linear light emitting sources L ₁ , L _2, ... Of the linear display 1 are, for example, 40
There are 0, so that a feature vector of 5 projection types consisting of 16 components for each character image to be recognized can be coded and displayed with 5 bits per component (5 projection types x 16 components x 5). bit).

In FIG. 1, the light emitted from the linear display 1 is collected by each cylindrical lens 2a of the cylindrical lens array 2 arranged on the front side along the longitudinal direction (orthogonal to the drawing) of each linear light emitting source L ₁ , L ₂ . It is illuminated and illuminated on the reference mask 3.

As shown in FIG. 4, the reference mask 3 has linear light sources L ₁ , L _2, ...
It has a two-dimensional spread in the longitudinal direction (X direction) and the arrangement direction (Y direction), and is composed of a code pattern composed of a transparent portion 3a and a non-transparent portion 3b shown by hatching. The horizontal code columns Y ₁ , Y ₂ ... Y _n represent one coding vector to be referred to, and in the above example, n is 400, which are linear light sources L ₁ , L _2.
There is a one-to-one correspondence with ... One coded vector represents one character to be referred to, and coded vector sequences X ₁ , X ₂ ... X _m corresponding to many character types are arrayed in the vertical direction of the mask 3. m is 500, for example.

In the example of FIG. 4, X ₁ has a distribution intensity value [90, 15, 75, ...] As a reference standard as its vector component, and the corresponding coding vector is the code string Y ₁ , Y ₂ ... Y _n . Translucent part 3a and non-translucent part 3b
It is represented by. Vector identification is achieved by matching the light-transmitting / non-light-transmitting pattern of one of X _{1 to} X _m with the lighting / non-lighting pattern of the linear light-emitting sources L ₁ , L ₂ ... .

In the example of FIG. ₁ , the lighting / non-lighting pattern of the linear light emitting sources L _{1 to} L ₅ is (1,1,1,1,0) = 75, and the corresponding reference mask 3
The translucent / non-transmissive pattern of (0,1,0,0,1) = 90 is shown. This is an example of a mismatch, the transmitted light is (0,1,0 ,,
0,0,) = 10. In the example of the reference mask 3 shown in FIG. ₄ , Y ₁ to Y ₅ of the vector X ₄ partially match the light emission patterns of the linear light emission sources L _{1 to} L ₅ shown in FIG.

The transmitted light of the reference mask 3 is received by the photosensor array 4 and is taken out as a binary electric signal. The photosensor array 4 has an array (n × m = 400 × 500) of two-dimensional light receiving elements 4a corresponding to the code string of the reference mask 3. Such an optical sensor array 4 can be composed of a CCD, a phototransistor array, a MOS-FET array, or the like.

FIG. 5 conceptually shows an electrical processing unit for vector extraction. The vector extractor comprises a block of a comparator 5, a voltage detector 6a and a switch 5b. The comparator 5 is composed of m (500) elements, and each element has m Y _i outputs read from the photosensor array 4 (parallel outputs of the photosensors corresponding to one column in FIG. 4). , And each belongs to the vector sequence X ₁ ~ X _m ) are given at the same time. On the other hand, the other inputs of the respective elements of the comparator 5 are commonly coupled, and the lighting patterns (1, 1, _{2) of the} linear light emitting sources L ₁ , L ₂ ...
0,0,1 ... etc.) corresponding to n code strings (y ₁ , y ₂ ... Y _i ...
y _n ) signals are sequentially given. Then, each time y ₁ → y ₂ → y ₃ changes, Y ₁ → Y ₂ → Y ₃ ...

Therefore, first, y _{1 of the} code string is compared with all Y ₁ belonging to X _{1 to} X _m . And X _i at which there is a match, i.e., the comparison output at the light that has passed through the matching mask 3 is detected positions X _i (s) is retained by the voltage detector 6a, the position light detection was not X is excluded by closing the switch 6b.
When this operation is further executed up to y _n and Y _{n in} order of y ₂ and Y ₂ , y ₃ and Y _3, ..., In principle, one vector X _x remains at the end. This vector X _x matches the projected feature vector of the input to be identified. When the lighting pattern y _i is 0 (non-lighting), the gate of the comparator 5 is closed and the comparison is not performed.

There may be a case where there are a plurality of match vectors that finally remain due to an error caused by the optical processing, but since the candidates are narrowed down, pattern matching may be performed by digital calculation processing (bit calculation).

FIG. 6 is a principle circuit example showing one element of the vector extractor shown in FIG. The comparator 5 can be constituted by an AND gate 5a, and the output Y _i of the sensor array 4 is given to one input thereof, and the lighting pattern y _i is given to the other input in synchronization with Y _i . When Y _i and y _i are 1 respectively, the AND gate output 1 outputs the latch circuit 6a via the AND gate 6a.
Is input to and latched. The initial value of the latch circuit 6a is 1, its output is returned to the AND gate 6a and the gate is opened. Therefore, as long as the comparison output of the AND gate 5a is 1, the content of the latch circuit 6a is 1. When the comparison output of the AND gate 5a becomes 0 (mismatch), the content held in the latch circuit 6a becomes 0, and the AND gate outputs the AND gate.
6b is closed. Therefore, thereafter, the content of the latch circuit 6a is fixed to 0, and the output thereof represents the vector X which is removed as the matching failure.

Finally, the output of the latch circuit 6a holding the value 1 indicates the vector X for which matching is established. When y _i is 0 (not lit), the latch circuit 6a is not triggered in order to invalidate the output of the AND gate 5a.

The vector identification device shown in FIG. 1 can be integrally manufactured by a semiconductor integrated device. The linear display 1 can be made of, for example, an LED array in which a large number of linear PN junction surfaces are arranged. The cylindrical lens array 2 is a flat plate lens in which a plurality of substantially semi-cylindrical lens portions whose refractive index changes in the radial direction and which are constant in the longitudinal direction are formed in a transparent plate such as glass or synthetic resin. Can be used. Further, a cylindrical gradient index lens or a glass rod in which glass rods having a uniform refractive index are arranged and fixed can be used. In short, it only has to have a function of collecting light in a linear shape. The reference mask 3 and the photosensor array 4 can be easily manufactured by well-known integrated circuit technology. In addition, the reference mask 3 can be used as a liquid crystal shutter,
The pattern of the reference mask 3 may be controlled by a standard code source such as a ROM by using a member capable of controlling non-light transmission.

Further, the element of the photosensor array may be configured by the light receiving part and the non-light receiving part (element lacking part) in association with the code string of the reference vector, and the photosensor array may have the function of the reference mask. .

[Effects of the Invention] As described above, the vector identification device of the present invention encodes the value of each component of the input vector and displays it on the linear display, and the reference vector is arranged in the direction orthogonal to the array direction of the display line elements. By a two-dimensional optical sensor array in which a large number of light-transmitting patterns or light-receiving patterns representing the code sequence of
Since the parallel collation is performed at the same time, it is possible to detect the output of the two-dimensional sensor array and electrically compare it with the code value of the input vector without performing a long-time digital processing operation. Can be discriminated and extracted. Therefore, a known input vector can be identified in a very short time.

In addition, errors that are likely to occur due to optical parallel processing are significantly reduced by the encoding method, and highly accurate matching and identification can be performed.

[Brief description of drawings]

1 is a sectional view of an essential part showing an embodiment of a vector identification device of the present invention, FIG. 2 is an intensity distribution vector diagram showing a projection feature, and FIG. 3 is a light emission pattern diagram of the linear display of FIG. FIG. 4 is a plan view of an essential part of the reference mask of FIG. 1, FIG. 5 is a block diagram of the principle circuit of the vector extractor, and FIG. 6 is a principle circuit diagram showing one element of the vector extractor. In the reference numerals used in the drawings, 1 ... Linear display 2 ... Cylindrical lens array 3 ... Reference mask 3a ... Translucent part 3b ... Non-transmissive part 4 ... Sensor array 5 ... Comparator 6a ... … Voltage detectors L ₁ and L ₂ …… Linear emission sources.

Claims (2)

[Claims] 1. A linear display that encodes the value of each component of an input vector and emits light as a sequence of a plurality of line elements, and a code sequence including a light-transmitting portion and a non-light-transmitting portion corresponding to each line element. A reference mask that extends in the array direction of the line elements to represent one reference vector, and the code strings of a plurality of different reference vectors are arrayed in a direction orthogonal to the array direction of the line elements; A two-dimensional photosensor array that receives the light from the linear display that faces the mask and passes through all of the light-transmitting parts, and outputs the output signal of the photosensor array and the code value of the input vector. A vector identification device characterized in that the input vector and one of the plurality of different reference vectors are identified based on the above. 2. A linear display that encodes the value of each component of an input vector and emits light as a row of a plurality of line elements, and a light receiving unit and a non-light receiving unit corresponding to each line element facing the linear display. A code string consisting of a part and a line extending in the array direction of the line elements to represent one reference vector, and the code strings of a plurality of different reference vectors are arrayed in a direction orthogonal to the array direction of the line elements. A two-dimensional optical sensor array, and based on the output signal of the optical sensor array and the code value of the input vector, the input vector and one of the plurality of different reference vectors are identified. A characteristic vector identification device.