JP4018089B2 - Method for generating data map for determining surface condition and determination method - Google Patents

Description

  The present invention relates to a technique for determining the surface state of an inspection object by image analysis.

  Inspection of scratches and other surface conditions on the surface of gears and other processed products is often performed by visual inspection. However, while the operation is simple, it requires concentration and is difficult. Therefore, there is a need for an apparatus that can inspect gears and other processed products without depending on visual observation.

  Japanese Patent Application Laid-Open No. 2004-228561 describes a method of determining the presence or absence of scratches such as dents on the gear tooth surface by photographing the gear tooth surface and evaluating the luminance of the obtained image. A gear tooth surface to be inspected and a gear tooth surface different from the tooth surface to be inspected at a predetermined interval with respect to one gear are created to create a plurality of images from substantially the same direction regarding the different tooth surfaces. A difference image representing a luminance difference is created. The difference image is added to create an added image, and the presence or absence of a dent is determined by comparing the luminance of the added image with a reference value.

  Japanese Patent Application Laid-Open No. 2004-228561 creates a self-organizing map from an image of an object using a neural network, forms a cluster of neurons corresponding to the same image, and authenticates the image of the object using this cluster Is described.

In Patent Document 3, a self-organizing map is first created from images of a plurality of inspection objects by using one pixel and luminance values of all pixels constituting the image as input data, and the self-organizing map is used as master data. A method is described in which the quality is determined based on the degree of matching between the inspection target image and the master data.
Patent Application Publication Number Sho 63-201556 Patent Application Publication No. 2002-109541 Patent application publication number 2003-44835

  In the technique described in Patent Document 1, a region that is not a scratch such as color unevenness or cutting marks on the surface of a gear or other processed product but has a luminance different from that of other image regions (hereinafter referred to as overdetection or overdetection region). Therefore, considerable visual work is required to detect scratches and other surface conditions. The present invention provides a technique capable of improving the accuracy of flaw detection by an automatic device and reducing the visual determination effort.

  The technique described in Patent Document 2 is authentication of an object, for example, a person, and is not suitable for determining the surface state of a workpiece such as a gear.

  In addition, the technique described in Patent Document 3 compares a part of an image to be inspected with master data and determines whether the inspection is good or not based on the difference. Master data is required for each inspection target, The precondition is that the shape is fixed. In the inspection of an object such as a gear surface, excessive detection (color unevenness and dust on the surface) regarded as a non-defective product has irregular and various shapes and does not satisfy such preconditions. Therefore, the method of Patent Document 3 cannot be applied to inspection of an object such as a gear.

  In one form (claim 1), this invention provides the method of producing | generating the data map for determining the surface state of the target object which has a plurality of the same structural units of the texture continuously. The method includes the steps of preparing a plurality of samples of an examination region including a determination required region, extracting a feature vector from the sample, inputting the feature vector into a self-organizing map, and learning. Including. The method further includes the steps of grouping adjacent neurons corresponding to each of the same sample in the learned self-organizing map into clusters, and each of the clusters having a predetermined surface, such as a scratch, a dent, and an exfoliation. Generating a determination data map by classifying it into one corresponding to a state and one not corresponding to a state.

  According to the present invention, a data map for automatically determining the surface state of an object having a plurality of structural units of the same texture (surface texture, pattern, etc.) of the same texture is automatically learned by a neural network. Therefore, a highly reliable determination data map can be obtained.

  In one embodiment of the present invention (claim 2), the clusters in the determination data map are classified into predetermined surface states and overdetection (surface unevenness, dirt, etc.).

  According to another aspect of the present invention (claim 3), the object is a gear, and the feature vector includes the area of the determination target area, the inclination of the area, the luminance value entropy of the inspection area, and the luminance value of the inspection area. One or more of anisotropy, average luminance of the edge image of the inspection area, and circularity of the determination required area are included. By generating a feature vector using these parameters, a highly reliable determination data map can be generated.

  According to another aspect (claim 4) of the present invention, there is provided a method for determining the surface state of an object using the data map. In this method, the target object is relatively rotated to sequentially capture the image of the structural unit, and the difference between the current image and the previous image is obtained in the image acquired from the imaging means; Detecting a determination required area having a luminance exceeding a predetermined threshold in the difference image generated by the difference means, and specifying an inspection area surrounding the area. The method further includes a step of extracting a feature vector from the examination region, a step of calculating a distance between the feature vector and a coupling coefficient vector of each neuron of the determination data map, and a neuron having a small distance adjacent to each other Determining a region, and determining a surface state based on the number of neurons belonging to a cluster of a predetermined surface state in the adjacent region.

  According to this method, the determination of the surface state is performed based on the distance calculation between the feature vector and the coupling coefficient vector of each neuron of the map using the determination data map generated by learning the sample of the determination required region. A highly reliable determination can be made.

  In an embodiment of the present invention (Claim 5), the step of determining the surface state is a ratio between the number of neurons belonging to a predetermined surface state cluster in the adjacent region and the total number of neurons in the adjacent region. When is equal to or greater than a predetermined value, the surface state is determined.

  In an embodiment of the present invention (Claim 6), the object is a gear, and the feature vector includes an area of a determination required region, an inclination of the region, a luminance value entropy of the inspection region, a luminance value anisotropy of the inspection region, One or more of the average luminance of the edge image of the inspection area and the circularity of the determination required area are included. By extracting feature vectors using these parameters, highly reliable surface state determination can be performed.

  In the embodiment of the present invention (Claim 7), the self-organization map is relearned using the determination result of the surface state, and the determination data map is updated. Thereby, inspection accuracy can be improved as use progresses.

  In another form (claims 8-10), the present invention is an apparatus using a computer.

  Next, an embodiment of the present invention will be described with reference to the drawings. FIG. 1 shows a configuration of an imaging apparatus when an object to be inspected is a gear 11. The gear 11 is attached to the rotary table 15 and rotated stepwise at a gear tooth pitch by a step motor (not shown) provided on the base 17. The gear 11 is irradiated by the lighting device 19. A camera 21 equipped with a CMOS image sensor photographs the tooth surface of the gear 11. Photographing is performed in synchronization with the stepwise rotation of the gear 11, and a still image is photographed for each tooth surface. In one embodiment, the lighting device 19 uses a blue light emitting diode, but a lamp or other lighting device can also be used. Further, by using a wide dynamic range CMOS camera having a wide measurable luminance range as the camera 21, a good image suitable for scratch determination can be obtained. Depending on the object, a camera using a CCD image sensor can also be used.

  The positioning sensor 18 is generally called a proximity sensor, and generates a high-frequency magnetic field by a detection coil. When a metal detection object approaches this magnetic field, an induced current (eddy current) flows through the detection object due to electromagnetic induction. This current changes the impedance of the detection coil and stops the high-frequency oscillation, thereby detecting the approach of the detection object. In this embodiment, the proximity of the gear crest is detected by this proximity sensor. A signal from the positioning sensor 18 is sent to the computer 23 (FIG. 2), and the timing unit 24 sends a signal to the camera 21 in response to the gear tooth detection. The camera 21 shoots the gear in response to this signal.

  FIG. 2 shows functional blocks of the computer 23 that processes an image photographed by the camera 21. The computer 23 may be a personal computer, a workstation or other computer. The image acquisition unit 25 receives an image captured for each tooth surface pitch from the camera 21 and passes it to the difference image generation unit 27. The difference image generation unit 27 calculates a difference between the image acquired this time and the image acquired last time, that is, the image of the gear tooth surface one pitch before. Since the tooth surfaces of the gear without any scratches or other surface conditions are substantially the same, taking a difference results in an image with a low overall brightness. If the gear tooth surface has a scratch such as a dent, the brightness of the scratched portion is different from that of the other portions, and therefore appears as a portion of high brightness in the difference image. When the difference between the two images is taken, there is a pixel having a negative luminance value, so that the intermediate luminance 128 is added to the luminance value of the difference to make the luminance values of all the pixels positive. A pixel whose luminance value exceeds 255 by adding 128 has a luminance value of 255, and a pixel that becomes negative even if 128 is added has a luminance value of 0. In this way, the luminance value range of the difference image is adjusted from 0 to 255.

  Referring to FIG. 4A, the inspection area setting unit 28 detects an area having a luminance higher than the threshold value from the difference image as a potential flaw area 31, and sets an area 33 surrounding the area as an inspection area. . The shape of the inspection region 33 may be a square or other polygons, or may be a circle or an ellipse. The size is set to a size sufficient to surround the wound area 31 according to the potential wound area 31.

  The feature extraction unit 29 extracts the following six parameters.

(1) Area S _{A of the} wound region 31
The inspection area is displayed in coordinates with the vertical axis being i and the horizontal axis being j, and the area S _A is obtained by integration using the coordinates of the boundary line of the scratch area 31.

(2) Long axis inclination θ _A with respect to the tooth streak when the wound region 31 is approximated to an ellipse
A dent often occurring in a gear can be approximated to an ellipse. In FIG. 4B, an approximate equation of an ellipse that approximates the flaw region 31 is obtained using a graphic approximation program. The approximate expression can be obtained using a known program. Typically, an approximate expression is obtained using a multivariate nonlinear least squares fitting program. An angle θ _A formed by the long axis 37 of the ellipse thus obtained and a straight line 35 parallel to the gear tooth stripe 13 (FIG. 1) is obtained.

(3) The luminance value entropy entB of the inspection area 33
The entropy of the luminance value of the inspection area 33 including the potential flaw area 31 is obtained by the following equation. The luminance value entropy represents the information amount of the luminance value distribution, and a large information amount represents that the degree of randomness of the luminance value distribution is large.

  In this embodiment, the randomness of the luminance value distribution from luminance values 0 to 255 is used as the luminance value entropy. In Equation 1, p is the luminance value, and rel [p] is the frequency of the luminance value p.

  FIG. 6 shows an example of a luminance value histogram. In the case of a gear tooth surface having dents or other scratches, since the scratches appear as pixel areas having various brightness values in the difference image, the brightness value entropy is large. (A) is a luminance value histogram of an image in which pixels having various luminance values randomly exist. Brightness values are widely distributed around the intermediate brightness 128, and the entropy is large. (B) is a luminance value histogram of an image with small variations in luminance values, and has a small entropy.

(4) Luminance value anisotropy ansB of inspection area 33
The anisotropy of the luminance value of the inspection area 33 is expressed by the following formula. Anisotropy indicates the degree of symmetry of the luminance value distribution.

  In this expression, k is a luminance value with the lowest appearance frequency. That is, ansB is a value obtained by dividing the luminance value entropy from 0 to the minimum luminance value in the inspection region 33 by 0 to the entropy of all luminance values. When the luminance value distribution is symmetrical on the left and right of the intermediate luminance value 125, ansB = −0.5. The closer ansB is to -0.5, the higher the symmetry and the smaller the anisotropy. The closer ansB is to -0 or -1 from -0.5, the lower the symmetry and the greater the anisotropy.

  Referring to FIG. 7, the luminance value histogram as shown in FIG. 7A has a high luminance symmetry and a small anisotropy because the luminance values are distributed without any bias around the intermediate value 128. In the luminance value histogram of FIG. 7B, the distribution of luminance values is different between the left side and the right side of the intermediate value 128, the symmetry is low, and the anisotropy is large. Assuming that the luminance value with the lowest appearance frequency is 230 in (B), ansB is close to -1.

(5) Circularity C _A of potential scratch area 31
Circularity C _A wound area 31 is expressed by the following equation.

In this equation, S _A is the area of the wound area 31, r _max is the maximum distance from the center of gravity of the wound area 31 to the end of the region 33.

(6) The average luminance edgB of the edge image in the inspection area 33
If represents the area of the inspection area 33 in S _B, the average luminance edgB edge images of the inspection area 33 is obtained by the following expression.

    Here, i and j indicate the coordinates of the pixel, and P (i, j) indicates the luminance value of the pixel at the coordinates i and j. ΔP (i, j) is the square of the difference between the luminance values of two pixels adjacent to the upper and lower sides (i-axis direction) of P (i, j) and the luminance of the two pixels adjacent to the left and right (j-axis direction) This is the square root of the sum of the difference squared. ΔP (i, j) represents the magnitude of the difference between the luminance values of adjacent pixels. Taking the difference between the luminance values of adjacent pixels corresponds to generating an image in which an edge portion having a large change in luminance value in the image is emphasized, that is, an edge image. edgB is obtained by dividing the sum of the magnitudes of the luminance values of adjacent pixels by the area of the region 33, and indicates the average luminance of the edge image.

  For each of the six parameters thus obtained, normalization is performed by the following equation using the maximum value and the minimum value for many samples.

Normalization = (obtained value-minimum value) / (maximum value-minimum value)
By this normalization, each parameter takes a value in the range of 0 to 1. Thus, the inspection region 33 including the potential flaw region 31 can be represented by a numerical vector of six feature parameters.

FIG. 5 shows the feature vector thus obtained. FIG. 5A shows a feature vector corresponding to a dent as seen in the image shown at the top of the figure. FIG. 5B shows a feature vector of color unevenness (overdetection) of the gear tooth surface. There is no significant difference between the area parameter S _A of the potential flaw area, the long axis inclination θ _A when the ellipse is approximated, and the anisotropy ansB, but the luminance value entropy entB, circularity, and the average of the edge image It can be seen that there is a large difference in luminance edgB.

  Returning to FIG. 2, the feature extraction unit 29 extracts the feature vector including the above-described six parameters from the inspection region 33 as described above. In an alternative embodiment, the feature extraction unit 29 extracts a feature vector composed of three parameters: the luminance value entropy of the inspection area, the circularity of the potential flaw area, and the average luminance of the edge image of the inspection area.

  FIG. 3 is a flowchart showing the flow of the flaw determination process for one gear tooth surface in one embodiment of the present invention. This process is continuously executed until the flaw determination for all tooth surfaces of the gear is completed. The The gear 11 is attached to the rotary table of FIG. In step S101, it is determined whether all gear tooth surfaces of the gear 11 have been inspected. If not completed, the process proceeds to step S103. To determine whether all gear tooth surfaces have been inspected, the number of teeth of the gear 11 is set in the counter, and the counter is decremented each time processing for one tooth is completed, and the value of this counter is checked. By doing.

  In step S103, the gear 11 is rotated by one tooth. The positioning sensor 18 detects the gear teeth and sends a detection signal to the computer 23. The timing unit 24 of the computer 23 sends a drive signal to the camera 21 in accordance with a signal from the positioning sensor 18. In response to this, the camera 21 captures the gear 11 in step S105 and sends the image to the image processing personal computer 23. When the number of shots is 1 (S107), that is, when the first tooth of the gear 11 has just been shot, image processing using the difference image cannot be performed. In order to take a picture, the subsequent steps are skipped, the first process is completed, and the process starts again from the start.

  In step S109, as described above, a difference image is generated from two consecutive tooth images, and an inspection region 33 surrounding the potential wound region 31 (FIG. 4) is selected. In step S111, a feature vector is extracted from the difference image as described with reference to FIG. 5, and is passed to step S113. In step S113, a scratch or other surface state is determined using the determination data map.

  First, a method for generating a determination data map by self-organizing learning using a neural network will be described.

In this embodiment, 40 × 40 neurons are used to generate a self-organizing map. FIG. 8 shows three neurons N _j−1 , N _j , N _{j + 1} . ω _j = (ω _j0 , ω _j1 ,..., ω _jm ) is a coupling coefficient vector of the neuron N _j , and χ = (χ ₀ , χ ₁ ,..., χ _m ) is an input vector is there. The self-organizing map is also called a self-organizing feature map and is an unsupervised algorithm. This map extracts hidden features in the input data and automatically learns to generate a map that selectively reacts to the input data. The self-organizing map was proposed by Kohonen and explained in, for example, “Neuro Fuzzy Genetic Algorithm” published in 1994 by Sangyo Tosho Co., Ltd.

  The input vector in this embodiment is a feature vector of the examination region 33 extracted by the feature extraction unit 29 (FIG. 2). A self-organizing map is created in the next step.

Step 1: A network initialization neuron connection vector ω _j = (ω _j0 , ω _j1 ,..., Ω _jm ) is determined for all neurons using random numbers.

Step 2: Input Vector Input Input vector χ = (χ ₀ , χ ₁ ,..., Χ _m ) is given to each neuron. The relationship between each neuron and the input vector is as shown in FIG.

Step 3: Distance calculation between neuron coupling coefficient vector and input vector The distance between each neuron coupling vector and the input vector is calculated by the following equation.

Step 4: Select a neuron with the smallest winner neuron determination distance d _j . This neuron is called the winner neuron and is represented by j *.

Step 5: Update the coupling coefficient (weight) of each of the learning winner neurons of the coupling coefficient vector and the neighboring neurons according to the following equation.

η is a positive constant and is 0.05 in this embodiment. h (j, j *) is called a neighborhood function and is represented by the following equation.

  σ (t) decreases as the learning progresses. Therefore, the range of the neighborhood function is wide at the initial stage of learning as indicated by a circle in FIG. In other words, the coarse adjustment changes to the fine adjustment as the learning progresses. Thus, the neighborhood function effectively generates a map. FIG. 9 shows the winner neuron as a small circle 91 and the neighborhood function surrounding it is represented as a large circle 93.

Step 6: Update t, return to input vector input The learning count t is updated to t + 1, return to the input vector input in Step 2, repeat Steps 2-6, and repeat the update of the coupling coefficient vector.

  In the self-organizing map, the winner neuron and nearby neurons all approach the current input vector. At the beginning of learning, many neurons are considered to be close by the neighborhood function, and a rough map is formed. As the learning progresses, the number of neurons considered to be close to the winner neuron by the neighborhood function decreases. For this reason, local fine adjustment progresses and spatial resolution increases.

Creation of judgment data map All the feature vectors of the inspection area 33 including scratches or overdetection obtained for multiple samples were selected in order or randomly and entered into the self-organizing map for learning. A self-organizing map reflecting the similarity relationship of feature vectors can be generated. That is, a cluster having similar features gathers on the self-organizing map. FIG. 10 shows the self-organizing map obtained in this way. In the example, 40 by 40 neurons were used, but in FIG. 10, it is shown as a 20 by 20 map for simplicity. In FIG. 10, each grid corresponds to one neuron. Blocks (1), (2),... (28) separated by solid lines represent clusters determined as follows.

  An image of the examination region having a feature vector having the smallest distance from the coupling coefficient vector is arranged at each neuron position of the self-organizing map. This examination area is one of a plurality of samples used for generating the self-organizing map. Adjacent neurons with the same image are selected and grouped. A group of neurons formed in this way is called a cluster. The grouping here is the same as the grouping function used in the existing drawing program, and the grouped neurons can be selected as a whole group and the properties can be set.

  Next, the images arranged in the respective clusters are visually confirmed, and it is determined whether the inspection region 33 includes a scratch or other surface condition or an overdetection such as surface unevenness. This determination result is set as a cluster property. Specifically, with “Overdetection” entered in the property fields of all clusters as the initial state, a self-organizing map is displayed on the computer display, and the clusters determined to be scratched are displayed with the mouse. Right-click and change and enter “Scratch”.

Determination view of a wound or overdetection 11 thus shows a map the cluster is classified. In this embodiment, the determination data map formed in this way is called a flaw data map. In FIG. 11, the hatched clusters (1), (2), (6), (10), (11), (12), (13), (15) and (16) blocks 55A, (22 ), (27) and the block 55B of (28) are clusters classified as “scratches”, and clusters without hatching are “overdetected” clusters. This flaw data map is registered in the computer 23 (FIG. 2), and the determination unit 30 uses it for determination of a flaw.

  Specifically, first, the distance between the feature vector passed from the feature extraction unit 29 and the coupling coefficient vector is obtained. This distance is calculated according to Equation 5 shown above. Neurons are selected in ascending order of the distance, a predetermined number, for example, 10 or less neurons are selected, and a circle or ellipse surrounding them is set as a neighborhood region. As is apparent from the generation process of the wound data map, neurons with approximate connection coefficient vectors are grouped as clusters, so if you select neurons in order of increasing distance as described above, a cluster of adjacent neurons is formed. Selected. A circle or ellipse surrounding this cluster of neurons is defined as a neighborhood region. For example, referring to FIG. 12, a circle 57 defines the neighborhood region.

  A circle 57 obtains the center position of the cluster of these neurons from the array on the wound data map of the cluster of selected neurons, and a circle whose radius is the distance from this center to the farthest neuron of the cluster of neurons. Can be obtained as

  Based on the ratio (K) / (S) of the number of neurons (K) belonging to the wound cluster included in the neighboring area defined by the circle or ellipse thus selected and the total number of neurons (S) in the neighboring area Check for scratches. When (K) / (S) ≧ D, it is judged as a scratch, and when (K) / (S) <D, it is judged as overdetection. D is a preset threshold value, for example, 0.5. In this way, the scratch and other surface conditions and overdetection are discriminated. In the example of FIG. 12, since there are eight neurons belonging to the wound cluster (12) in the circle 57 and the total number of neurons in the circle 57 is 12, (K) / (S) = 8/12 = Since it is 0.67, it is determined that a scratch is included in the current examination area.

Update of self-organizing map The tooth surface of the gear determined to be flawed by automatic determination by a computer using the flaw data map described above is visually inspected to confirm the flaw. Similarly, the tooth surface of the gear determined to be overdetected by automatic determination can be visually inspected and confirmed. Thus, the feature vector of the inspection region of the gear tooth surface finally determined to have a scratch or other surface condition is input to the self-organization map before clustering to be learned. At this time, the initial value of the coupling coefficient vector of the neurons of the self-organizing map uses the latest coupling coefficient vector generated by past learning.

  In this way, the self-organizing map is updated, and the above-described clustering process is executed again. As a result, a cluster corresponding to a new type of flaw or overdetection can be generated and a flaw data map can be generated without reducing the detection accuracy of existing types of flaws. By using the updated flaw data map, the inspection accuracy can be improved. In this update process, since the coupling coefficient vector obtained in the past is used, the flaw data map can be updated in a short learning time.

  Although the present invention has been described with respect to specific embodiments, the present invention is not limited to such embodiments.

The block diagram which shows the whole structure of the imaging device of a gear tooth surface. The functional block diagram which shows the whole structure of an image processing computer. The flowchart which shows the flow of the process which determines the damage | wound of a gear tooth surface. The figure which shows the relationship between a potential flaw area | region and a test | inspection area | region. The figure which shows a feature vector. The figure which shows the luminance value histogram from which entropy differs. The figure which shows the luminance value histogram from which anisotropy differs. The figure which shows the partial structure of a neural network. The figure which shows the production | generation process of a self-organization map. The figure which shows the self-organization map divided into the clusters. The figure which shows the damage | wound data map produced | generated from the self-organization map. The figure for demonstrating the damage determination using a damage | wound data map.

Explanation of symbols

11 Gear 19 Lighting device (LED)
21 Camera 23 Computer

Claims (9)

A method for generating a data map for determining a surface state of an inspection object having a plurality of structural units of similar texture continuously,
Preparing a plurality of samples of an inspection area including a determination required area;
Extracting a feature vector from the sample;
Inputting the feature vector into a self-organizing map and learning;
Grouping adjacent neurons corresponding to each of the same samples in the learned self-organizing map into clusters;
Classifying the cluster as a flaw or overdetection according to a determination result input by a user as to whether the cluster corresponds to a flaw or overdetection ;
A method for generating a data map having:   The object is a gear, and the feature vector includes an area of the determination target region, an inclination of the region, a luminance value entropy of the inspection region, a luminance value anisotropy of the inspection region, and an average luminance of an edge image of the inspection region. And the degree of circularity of the determination required area. A determination method using the data map according to claim 1,
Relatively rotating the object and sequentially capturing images of the structural units;
In the image acquired from the imaging means, taking a difference between the current image and the previous image;
Detecting a required determination area having a luminance exceeding a predetermined threshold in the difference image generated by the difference means, and specifying an inspection area surrounding the required determination area;
Extracting a feature vector from the inspection region;
Calculating a distance between the feature vector and a coupling coefficient vector of each neuron of the data map;
Obtaining a region adjacent to a neuron having a small distance; and
Determining a surface state based on the number of neurons belonging to a cluster of flaws in the adjacent region;
A method for determining a surface state, including: The step of performing the determination includes determining a surface state when a ratio between the number of neurons belonging to the wound cluster in the adjacent region and the total number of neurons in the adjacent region is equal to or greater than a predetermined value. Item 4. The determination method according to Item 3.   The object is a gear, and the feature vector includes an area of the determination required region, an inclination of the region, a luminance value entropy of the inspection region, a luminance value anisotropy of the inspection region, and an average luminance of an edge image of the inspection region. And the degree of circularity of the determination-needed region.   The feature vector of the inspection region of the object to be determined is input to the self-organizing map to be learned to update the self-organizing map, and based on the updated self-organizing map The method of claim 3, comprising updating the data map. An apparatus using a computer that generates a data map for determining a surface state of an object having a plurality of similar texture units continuously.
Means for extracting a feature vector from a plurality of samples in an inspection area including a determination required area;
A neural network that performs self-organization learning using the feature vector as input and generates a self-organization map;
Means for grouping and clustering adjacent neurons corresponding to each of the same samples in the learned self-organizing map;
Means for classifying the cluster into a flaw or overdetection according to a determination result input by a user as to whether the cluster corresponds to a flaw or overdetection ;
A device for generating a data map having: A determination apparatus using the data map according to claim 7,
Means for relatively rotating the object and sequentially capturing images of the structural units;
In the image acquired from the imaging means, means for taking the difference between the current image and the previous image;
Means for detecting a determination required area of luminance exceeding a predetermined threshold in the difference image generated by the difference means, and specifying an inspection area surrounding the area;
Means for extracting a feature vector from the inspection region;
Means for calculating a distance between the feature vector and a coupling coefficient vector of each neuron of the data map;
And a means for obtaining a region adjacent to a neuron with a small distance,
A surface state determination device that determines a surface state based on the number of neurons belonging to a cluster of flaws in the adjacent region.   The object is a gear, and the feature vector is an area of the determination target region, an inclination of the region, a luminance value entropy of the inspection region, a luminance value anisotropy of the inspection region, an average luminance of an edge image of the inspection region, The apparatus according to claim 8, comprising one or more of: and a circularity of the determination required area.

Images