JP6549396B2 - Region detection apparatus and region detection method - Google Patents

Description

  The present invention relates to an area detection device and an area detection method.

  2. Description of the Related Art In the appearance inspection of a substrate such as a semiconductor substrate, a glass substrate, and a printed wiring board, automatic defect classification has been conventionally performed to automatically classify an image of a detected defect. In automatic defect classification, a discriminant function is estimated using feature quantities calculated from defect images. Higher-order local auto-correlation (HLAC) feature quantities are known as one of such feature quantities.

  Patent Documents 1 and 2 disclose an adaptive learning type general-purpose image measurement method using HLAC for a two-dimensional image. HLAC satisfies position invariance and additivity, which are fundamental requirements in image recognition, and is characterized by low computational complexity and high recognition performance. If the degree of correlation in the HLAC is limited to the second order, there are 35 mask patterns (displacement patterns) for the multi-valued image when the displacement is near 8 and 25 for the binary image. There is a mask pattern. By adding the values obtained by the calculation according to the mask pattern to each pixel in the entire image, the HLAC feature value by the mask pattern can be obtained.

  Patent Document 3 discloses a method of generating a binary image for classification. In this method, a first threshold in which the shape of a defect is detected in a difference image between a captured image and a reference image, and a second threshold larger than the first threshold and in which detection of background noise is reduced are prepared. The difference image is binarized with a first threshold to generate a first binary image, and the difference image is binarized with a second threshold to generate a second binary image. A plurality of connected regions are acquired by the labeling process for the first binary image, and among the plurality of connected regions, connected regions not including the pixel at the position detected in the second binary image are specified and deleted. Thereby, a classification binary image is generated.

Patent No. 2982814 Patent No. 2834153 JP, 2013-134666, A

  By the way, in visual inspection of a substrate, in order to classify a defect with high accuracy, it is necessary to appropriately detect a defect area. Generally, when a target area such as a defect area is detected from an image obtained by capturing an object, the captured image is compared with a reference image indicating an object without abnormality. However, in this case, it is necessary to prepare a reference image for all of the areas to be inspected, and detection of the target area becomes complicated. There are also cases where preparation of reference images is difficult.

  The present invention has been made in view of the above problems, and aims to easily detect a target area.

The invention according to claim 1 is an area detection device for detecting a target area used for a predetermined purpose from a target image, wherein a plurality of mask patterns different from one another are prepared, and each pixel of the target image is centered. A plurality of pixel features respectively corresponding to the plurality of mask patterns are arranged by arranging each mask pattern smaller in size than the window at a plurality of positions and performing an operation according to each mask pattern within the window A pixel feature amount calculating unit which obtains an amount for each of the pixels, a pixel classifier which determines whether each of the pixels is included in a target area based on the plurality of pixel feature amounts, and the pixel classification a region feature calculation unit for calculating a region feature for the detected the object regions by vessel, a region classifier for classifying the object region based on the region feature Wherein the region feature calculation portion, the values obtained by calculation in accordance with the mask pattern with respect to each pixel included in the target area as the pixel calculation value, the pixel calculation values of all the pixels included in the target area the sum of, you calculated as the region feature.

  The invention described in claim 2 is the area detection device according to claim 1, wherein the plurality of mask patterns include mask patterns for high-order local autocorrelation feature quantities.

The invention according to claim 3 is the area detection device according to claim 1 or 2 , wherein the area feature quantity calculation unit makes a minute area with respect to an image which distinguishes the target area from other areas. Preprocessing for removal is performed, and the region feature amount is calculated for the target region indicated by the preprocessed image.

The invention according to claim 4 is the area detection apparatus according to any one of claims 1 to 3 , wherein the target image is an image obtained by imaging an object, and the target area is on the object. Is an area showing a defect of

The invention according to claim 5 is the area detection device according to claim 4 , wherein a defect area image obtained based on a difference image between a defect image indicating a defect and a predetermined reference image, and the defect image The pixel classifier is constructed by performing learning using a plurality of pixel feature quantities for each pixel of.

The invention according to claim 6 is an area detection method for detecting a target area used for a predetermined purpose from a target image, wherein a) a plurality of mask patterns different from one another are prepared, and each pixel of the target image By arranging each mask pattern smaller in size than the window at a plurality of positions and performing an operation according to each mask pattern in a window centered on the plurality of mask patterns respectively corresponding to the plurality of mask patterns. Acquiring a pixel feature amount for each pixel; b) determining whether each pixel is included in a target area based on the plurality of pixel feature amounts ; c) b) comprising calculating a region feature for the detected the object regions in step, and a step of classifying the region of interest based on d) the region feature, said step c) In addition, the sum of the pixel operation values of all the pixels included in the target area is a feature of the area, with the value obtained by the operation according to the mask pattern for each pixel included in the target area as a pixel operation value. Ru is calculated as the amount.

The invention according to claim 7 is the region detection method according to claim 6 , wherein the plurality of mask patterns include a mask pattern for high-order local autocorrelation feature value.

The invention according to claim 8 is the region detection method according to claim 6 or 7 , wherein the step c) removes a minute region with respect to an image showing the target region distinguished from other regions. The method includes the steps of performing preprocessing, and calculating the region feature amount for the target region indicated by the preprocessed image.

The invention according to claim 9 is the area detection method according to any one of claims 6 to 8 , wherein the target image is an image obtained by imaging an object, and the target area is on the object. Is an area showing a defect of

The invention according to claim 10 is the area detection method according to claim 9 , wherein a defect area image obtained based on a difference image between a defect image indicating a defect and a predetermined reference image, and the defect image The pixel classifier used in the step b) is constructed by performing learning using a plurality of pixel feature amounts for each of the pixels.

According to the present invention, a target area can be easily detected. In addition, the target area can be classified with high accuracy.

It is a figure showing composition of an inspection / classification device. It is a figure showing composition of a computer. It is a block diagram showing functional composition in an inspection / classification device. It is a figure which shows the flow of construction of a pixel classifier and an area | region classifier. It is a figure which shows a teacher image set. It is a figure which shows a teacher image set. It is a figure which shows a teacher image set. It is a figure for demonstrating the process which calculates a pixel feature-value. It is a figure which shows a detection area image and a processed detection area image. It is a figure which shows a detection area image and a processed detection area image. It is a figure which shows a detection area image and a processed detection area image. It is a figure showing a flow of inspection and classification by an inspection and classification device. It is a figure showing an object picture. It is a figure which shows a detection area image and a processed detection area image. It is a figure which shows a defect area | region image. It is a figure which shows a defect image. It is a figure which shows a reference image. It is a figure which shows a defect area | region image. It is a figure showing an object picture. It is a figure which shows a detection area image.

  FIG. 1 is a view showing the configuration of an inspection and classification device 1 according to an embodiment of the present invention. The inspection / classification apparatus 1 is an apparatus for detecting a defect of a pattern on a semiconductor substrate 9 (hereinafter simply referred to as “substrate 9”) and classifying the defect. The inspection / classification apparatus 1 controls the entire operation of the inspection / classification apparatus 1 and the computer 5 that realizes an inspection section, a classification section, and a learning section described later. Equipped with The computer 5 performs a defect inspection based on the image data from the imaging unit 2, and when a defect is detected, the defect is classified into a defect class (which is a type of defect and is also referred to as a "category"). A process of automatically classifying defects is performed. The classes of defects present on the substrate 9 are, for example, chips, protrusions, breaks, shorts, foreign matter, and the like. The inspection and classification apparatus 1 can also be regarded as an apparatus in which the function of automatic defect classification is added to the defect inspection apparatus.

  The imaging unit 2 captures an inspection target area on the substrate 9 to obtain (a data of) a multivalued captured image, a stage 22 for holding the substrate 9, and a stage 22 for the imaging device 21. And the stage drive unit 23 that moves relative to each other. The imaging device 21 converts the image of the substrate 9 imaged by the optical system 212 into an illumination unit 211 that emits illumination light, an optical system 212 that guides the illumination light to the substrate 9 and the light from the substrate 9 is incident. It has an area sensor 213 that converts it into a signal. The stage drive unit 23 is configured by a ball screw, a guide rail, a motor, and the like, and the computer 5 controls the stage drive unit 23 and the imaging device 21 to capture an image of the inspection target area on the substrate 9. In the imaging device 21, a line sensor in which light receiving elements are arranged in one dimension may be used instead of the area sensor 213. In this case, inspection is performed by moving the stage 22 in a direction perpendicular to the arrangement direction of the light receiving elements. A two-dimensional captured image showing a target area is acquired.

  FIG. 2 is a diagram showing the configuration of the computer 5. The computer 5 has a general computer system configuration including a CPU 51 for performing various arithmetic processing, a ROM 52 for storing a basic program, and a RAM 53 for storing various information. The computer 5 includes a fixed disk 54 for storing information, a display 55 for displaying various information such as images, a keyboard 56a and a mouse 56b (hereinafter collectively referred to as "input unit 56") for receiving user input, and an optical disk. , A reader 57 for reading information from a computer readable recording medium 8 such as a magnetic disk or a magneto-optical disk, and a communication unit 58 for transmitting and receiving signals to and from other configurations of the inspection and classification device 1. Including.

  In the computer 5, the program 80 is read from the recording medium 8 through the reader 57 in advance and stored in the fixed disk 54. Then, arithmetic processing is performed according to the program 80 while using the RAM 53 and the fixed disk 54 by the CPU 51.

  FIG. 3 is a block diagram showing a functional configuration in the inspection / classification apparatus 1. In FIG. 3, the functional configuration realized by the CPU 51, the ROM 52, the RAM 53, the fixed disk 54, etc. It is enclosed by. The computer 5 uses an inspection unit 31 that detects a defect based on a captured image acquired by the imaging unit 2, and the defect using a neural network, a decision tree, a discriminant analysis, etc. when the defect is detected. It has classification part 32 which classifies automatically. The inspection unit 31 includes a pixel feature quantity calculation unit 311 and a pixel classifier 312. The classification unit 32 includes an area feature quantity calculation unit 321 and an area classifier 322. The computer 5 further includes a learning unit 33 for training the pixel classifier 312 and the area classifier 322. Details of functions implemented by these configurations will be described later. Note that these functions may be constructed by dedicated electric circuits, or partially dedicated electric circuits may be used.

  In the inspection / classification apparatus 1, the pixel classifier 312 and the region classifier 322 are constructed by the learning unit 33 as preparation in advance. FIG. 4 is a diagram showing a flow of construction of the pixel classifier 312 and the area classifier 322. Constructing a classifier means generating a classifier by assigning values to parameters included in the classifier, determining a structure, and the like.

  At the time of construction of a classifier, first, a plurality of captured images (hereinafter referred to as "defect image") showing various defects are prepared. In addition, a difference image (typically, an image showing the absolute value of the difference between both images) is acquired by comparing each defect image with a reference image having no defect, and the difference image is binarized, etc. Thus, an image (hereinafter referred to as a "defect area image") indicating the area of the defect is generated. In the defect area image, for example, the value 1 is assigned to the pixels included in the area of the defect and the value 0 is assigned to the pixels not included in the area of the defect. In the defect area image, a process of expanding the area of the defect may be appropriately performed. In the generation of the defect area image, another method such as the method of Japanese Patent Application Laid-Open No. 2013-134666 (Patent Document 3) may be used, for example.

  The plurality of defect images are displayed on the display 55 of the computer 5. When the input unit 56 receives an input from the user, one defect class out of the plurality of defect classes is associated with each of the plurality of defect images. As a result of the above processing, a plurality of teacher image sets are prepared with a set of defect images taught to belong to one defect class and defect area images corresponding to the defect images as a teacher image set (step S11). . In the following description, the defect class assigned to each defect image by the user is referred to as a "teaching defect class".

  5A-5C are diagrams showing a set of teacher images. 5A to 5C show the defect image 71 and the defect area image 72. The pixel feature quantity calculation unit 311 acquires a pixel feature quantity vector for each pixel in the defect image 71 of each teacher image set (step S12). In the present embodiment, higher-order Local Auto-Correlation (HLAC) in an area of a predetermined size centered on each pixel of the defect image 71 (an area in the window 6 of FIG. 6 described later). The feature amount is obtained as a pixel feature amount.

  FIG. 6 is a diagram for explaining the process of calculating the pixel feature amount of each pixel of the defect image 71. In the defect image 71 in which a plurality of pixels 711 are arranged in the row direction and the column direction, assuming that one pixel 711a to which parallel diagonal lines are hatched in FIG. In the image 71, a window 6 centered on the pixel of interest 711a is set. For example, the length in the row direction of the window 6 in FIG. 6 is nine times the pixel pitch P1 in the row direction, and the length in the column direction of the window 6 is nine times the pixel pitch P2 in the column direction.

  Further, in the pixel feature quantity calculation unit 311, a plurality of mask patterns 61 for the HLAC feature quantity are prepared. The plurality of mask patterns 61 are different from one another. In each mask pattern 61 (also referred to as a kernel), a plurality of elements are arranged in the row direction and the column direction, and each element has the same size as the pixel 711. For example, the length in the row direction of each mask pattern 61 is three times the pixel pitch P1 in the row direction, and the length in the column direction of the mask pattern 61 is three times the pixel pitch P2 in the column direction. The size of each mask pattern 61 is smaller than the size of the window 6. In FIG. 6, the thickness of the solid line indicating the outline of the mask pattern 61 is drawn thicker than the solid line indicating the outline of the window 6. The length of the window 6 in the row direction and the column direction is preferably at least twice and at most 5 times the length of the mask pattern 61.

  The mask pattern 61 scans the inside of the window 6 as indicated by an arrow with a reference A1 in FIG. That is, under the condition that the entire mask pattern 61 is disposed in the window 6, at each position in the window 6 in the column direction, the mask pattern 61 is one pixel from one end to the other end in the window 6 in the row direction. Move by Further, in the mask pattern 61, a plurality of elements to be multiplied with each other are specified, and at each position where the mask pattern 61 is arranged, values of a plurality of pixels 711 overlapping with a plurality of elements indicated by the mask pattern 61 are selected. Are multiplied to obtain an intermediate operation value. Then, the sum of intermediate operation values at all positions where the mask pattern 61 is arranged in the window 6 is obtained as a pixel feature value by the mask pattern 61. In the plurality of mask patterns 61, the elements to be multiplied are different from one another.

  As described above, by arranging each mask pattern 61 at a plurality of positions in the window 6 and performing an operation according to the mask pattern 61, a plurality of pixel feature amounts respectively corresponding to the plurality of mask patterns 61 are noticed It is acquired for the pixel 711a. Then, an array of a plurality of pixel feature amounts is acquired as a pixel feature amount vector for the target pixel 711a. The above process is performed with each pixel 711 in which the window 6 can be placed (that is, each of all the pixels 711 excluding the pixel 711 located near the outer edge of the defect image 71) in the defect image 71 as a pixel of interest Pixel feature value vectors for approximately all 71 pixels 71 are obtained. In other words, a plurality of pixel feature amount images respectively corresponding to a plurality of (for example, 35) mask patterns 61 are acquired. The size of the pixel feature amount image is slightly smaller than the defect image 71 according to the size of the window 6.

  On the other hand, the defect area image 72 of each teacher image set indicates whether each pixel 711 of the defect image 71 is included in the defect area indicating the defect on the substrate 9 or in the non-defect area not indicating the defect. That is, the defect area image 72 indicates the area class (defect area and non-defect area) to which each pixel 711 of the defect image 71 belongs. The learning unit 33 generates a pixel classifier 312 using the pixel feature quantity vector of each pixel 711 of the defect image 71 included in the plurality of teacher image sets and the area class of the pixel 711 indicated by the defect area image 72. (Step S13). The combination of the pixel feature amount vector of each pixel 711 and the area class of the pixel 711 can be regarded as teaching data for constructing the pixel classifier 312. As a method or mechanism for generating the pixel classifier 312, various learning type (teaching type) may be used. For example, the nearest neighbor method, the minimum distance method, the subspace method, the decision tree method, the linear discriminant method, the fuzzy voting method, the flexible nib Bayesian method, the neural network, the support vector machine (SVM), etc. can be adopted.

  Subsequently, the area feature quantity calculation unit 321 acquires an area feature quantity vector for each teacher image set (step S14). In the present embodiment, the HLAC feature value is determined in the defect area of the defect image 71 specified by the defect area image 72. That is, the sum of pixel operation values of all the pixels included in the defect area is a value obtained by performing an operation according to the mask pattern for each pixel included in the defect area of the defect image 71 as the pixel operation value. It is obtained as an area feature value by the mask pattern. Thus, the area for which the area feature value is to be obtained is limited to the defect area in the defect image 71.

  In the area feature quantity calculation unit 321, as in the pixel feature quantity calculation unit 311, a plurality of mask patterns are prepared, and the area feature quantity by each mask pattern is obtained. Then, an array of a plurality of region feature amounts obtained respectively using a plurality (all) of mask patterns is acquired as a region feature amount vector. The region feature may be obtained using the pixel feature calculated by the pixel feature calculator 311. For example, it may be the sum of pixel features of all pixels included in the defect region. . In addition to HLAC feature quantities, feature quantities such as defect area, perimeter length, barycentric position, moment quantity, average of tone values, variance, maximum, minimum, etc. are obtained from defect image 71 or defect area image 72. , And may be included in the region feature vector.

  When the region feature vector is obtained for each teacher image set, the region feature vector of the plurality of teacher image sets and the teaching defect class associated with the defect image 71 included in the plurality of teacher image sets are used. Then, the area classifier 322 is generated (step S15). The combination of the region feature vector of each defect region and the taught defect class of the defect region (that is, the taught defect class of the defect image 71) can be regarded as teaching data for constructing the region classifier 322. Similar to the generation of the pixel classifier 312, various methods or mechanisms for generating the region classifier 322 may be used as long as they are of a learning type (teaching type). The pixel classifier 312 and the area classifier 322 are constructed by the above processing.

  Here, when each pixel of the defect image 71 of the teacher image set shown in FIG. 5A to FIG. 5C is classified by the pixel classifier 312 constructed by the process of FIG. , “Detected area image 73” is obtained. The detection area image 73 is a binary image in which the value of the pixel determined to belong to the defective area by the pixel classifier 312 is different from the value of the pixel determined to belong to the non-defective area. In other words, the detection area image 73 is an area in which the area class is a defect area (that is, a detected defect area, hereinafter referred to as a “detection defect area”), and another in which the area class is a non-defect area. It is an image shown separately from the area.

  The white detection defect area in the detection area image 73 of FIGS. 7A to 7C substantially includes the defect area (white area) indicated by the defect area image 72 of FIGS. 5A to 5C, but a minute detection defect area not corresponding to the defect area. (Including detection defect areas with a minute width.) Are also scattered. As described later, in the inspection and classification in the inspection and classification apparatus 1, when the inspection defection area is acquired by the inspection part 31 from the captured image in which the presence or absence and the position of the defect area are unknown, the classification part 32 detects Classification of defect areas is performed. In the present embodiment, since the minute detection defect area is highly likely to be a false defect, in the classification of the defect class in the classification unit 32, it is preferable that the minute detection defect area is removed.

  In this case, in the region feature quantity calculation unit 321 in the classification unit 32, the detection region image 73 is subjected to preprocessing (noise removal processing) for removing a minute region (including a region with a minute width), An area feature amount vector is acquired for a detected defect area indicated by the processed detection area image. In the present embodiment, contraction processing and expansion processing on a detection defect area of the detection area image 73 are performed as preprocessing. In the right side of FIGS. 7A to 7C, an image 74 obtained by performing contraction processing on the detection defect area of the detection area image 73 on the left side and subsequently performing expansion processing similar to the contraction processing (preprocessing Hereinafter, it is referred to as “processed detection area image 74”. The processed detection area image 74 on the right side of FIGS. 7A to 7C is closer to the defect area image 72 on the right side of FIGS. 5A to 5C than the detection area image 73 on the left side.

  FIG. 8 is a diagram showing the flow of inspection and classification by the inspection and classification device 1. First, as the imaging unit 2 images the inspection target area of the substrate 9, a multivalued captured image is acquired and input to the inspection unit 31 as a target image (step S21). FIG. 9 shows a part of the target image 70, and in practice, the target image 70 showing a wide area on the substrate 9 is obtained. In the target image 70, the presence and the position of the defect area are unknown. The images in FIG. 10 and FIG. 11 described later correspond to FIG. In practice, a plurality of inspection target areas on the substrate 9 are sequentially imaged by the imaging unit 2 to acquire a plurality of target images 70, and the following processing is sequentially performed on the plurality of target images 70.

  The pixel feature quantity calculation unit 311 obtains a pixel feature quantity vector for each pixel of the target image 70 by the same process as step S12 in FIG. 4 (step S22). That is, in the window 6 (see FIG. 6) centered on each pixel of the target image 70, the mask patterns 61 are arranged at a plurality of positions, and calculations are performed according to the mask patterns 61 to obtain a plurality of masks. A plurality of pixel feature quantities respectively corresponding to the pattern 61 are acquired. The pixel classifier 312 determines whether the pixel is included in the defect area based on the pixel feature quantity vector (step S23). For example, the pixel determined to be included in the defect area (hereinafter referred to as "detected defect pixel") is assigned the value 1 and the pixel determined not to be included in the defect area is assigned the value 0. Be done. Thus, a detection area image 73 shown on the left side of FIG. 10 is obtained with respect to the target image 70 shown in FIG. The detection area image 73 shows a detection defect area (white area). Each detection defect area is a set (closed area) of detection defect pixels adjacent to each other, and is also called a blob.

  The detection area image 73 is input to the classification unit 32. In the region feature quantity calculation unit 321, a contraction process is performed on each detection defect area of the detection area image 73, and then, an expansion process similar to the contraction process is performed. Thus, preprocessing (noise removal processing) for removing a minute area is performed on the detection area image 73, and a processed detection area image 74 shown on the right side of FIG. 10 is obtained (step S24). Note that the preprocessing in the region feature quantity calculation unit 321 may be performed by another method as long as a minute region in the detection region image 73 can be removed.

  The area feature quantity calculation unit 321 acquires an area feature quantity vector for each detection defect area of the processed detection area image 74 by the same process as step S14 in FIG. 4 (step S25). For example, in the detection defect area of the target image 70 specified by the processed detection area image 74, the HLAC feature value is obtained. In addition to the HLAC feature quantities, feature quantities such as defect area, perimeter, center of gravity position, moment quantity, average of tone values, variance, maximum, minimum, etc. are acquired from the target image 70 or the processed detection area image 74 , And may be included in the region feature vector. The area feature value vector for each detection defect area in the processed detection area image 74 is input to the area classifier 322, and the classification result is output. That is, the area classification unit 322 classifies the detected defect area (blob) into any one of a plurality of defect classes based on the area feature quantity vector (step S26). In the classification unit 32, calculation of the area feature amount is performed in real time each time the inspection unit 31 detects a detected defect area, and automatic classification of a large number of detected defect areas is performed at high speed.

  Note that FIG. 11 shows a defect area image 72 obtained by generating a difference image by comparing the target image 70 of FIG. 9 with a reference image and binarizing the difference image. The detection area image 73 on the left side of FIG. 10 and the processed detection area image 74 on the right side are similar to the defect area image 72 of FIG. Thus, in the detection area image 73 and the processed detection area image 74, the defect area is appropriately detected, and the detection area image 73 and the processed detection area image 74 are replaced with the defect area image 72 to classify the defect area. It can be seen that it is available to

  As described above, in the inspection / classification apparatus 1 which is an area detection apparatus, each mask pattern 61 is arranged at a plurality of positions in the window 6 centered on each pixel of the target image 70 and the mask pattern 61 is arranged. A plurality of pixel feature amounts respectively corresponding to the plurality of mask patterns 61 are obtained for the pixels by performing the calculation according to the above. Then, based on the plurality of pixel feature amounts, it is determined whether the pixel is included in a defect area indicating a defect on the substrate 9 or not. Thereby, in the target image 70, it is realized to easily detect a detection defect area (defect area) which is a target area used for detection and classification of a defect without using a reference image.

  In addition, an area feature amount is calculated for the detected defect area, and the detected defect area is classified based on the area feature amount. Thereby, the detection defect area can be appropriately classified. By calculating the HLAC feature quantity for the detection defect area in the target image 70 as the area feature quantity, the detection defect area can be classified with high accuracy. Further, the detection area image 73 showing the detected defect area separately from the other areas is subjected to pre-processing for removing a minute area, and the area feature amount for the detected defect area indicated by the pre-processed detection area image 74 The detection defect area can be classified more appropriately by calculating.

  The inspection / classification apparatus 1 uses a defect area image 72 acquired based on a difference image between a defect image 71 indicating a defect and a predetermined reference image, and a plurality of pixel feature amounts for each pixel of the defect image 71. By performing learning, the pixel classifier 312 is constructed. This makes it possible to omit or simplify the teaching operation (teaching operation for the pixels included in the defect area) by the user in the construction of the pixel classifier 312.

  Various modifications can be made to the inspection / classification apparatus 1 and the area detection method.

  For example, in the window centered on each pixel of the target image (defective image), the density gradient direction of the plurality of positions can be determined by arranging the differential filter, which is a mask pattern, in the plurality of positions. A value indicating the dispersion in the density gradient direction of the position may be acquired as one pixel feature amount for the pixel. Thus, in the pixel feature quantity calculation unit 311, various mask patterns other than the mask pattern for the high-order local autocorrelation feature quantity may be used.

  The target image may show the appearance of a substrate other than the semiconductor substrate. That is, the defect image may indicate a defect of a pattern of a substrate other than the semiconductor substrate or a defect such as a foreign substance. Examples of the substrate include thin film devices such as hard disk substrates, glass substrates used for thin displays such as plasma displays and liquid crystal displays, photomask substrates, film substrates, printed wiring substrates and the like.

  Moreover, the inspection and classification device 1 may be used for inspection and classification of an image obtained by imaging a solar cell panel. In this case, for example, an image obtained by imaging EL (electro luminescence) or PL (photo luminescence) emission of a solar cell panel, or an image of a solar cell panel obtained using a laser terahertz emission microscope (LTEM) An area different from the normal area indicated by the reference image is the defect area (target area) to be detected.

  12 and 13 show images obtained by imaging the EL emission of the solar cell panel, FIG. 12 is a defect image showing a region including a defect, and FIG. 13 is a reference showing a region not including a defect It is an image. FIG. 14 shows a defect area image obtained based on the difference image between the defect image and the reference image. The pixel feature quantity calculation unit 311 of the inspection / classification apparatus 1 calculates a plurality of pixel feature quantities for each pixel of the defect image, and performs learning using the defect area image and the pixel feature quantity vector of each pixel. A pixel classifier 312 is constructed. Then, when the target image shown in FIG. 15 in which the presence or absence and the position of the defect area are unknown is input, it is determined whether the pixel is included in the defect area based on the pixel feature quantity vector of each pixel of the target image. And a detection area image shown in FIG. 16 is acquired. Of course, in the same manner as the above embodiment, the region classifier 322 may be constructed, and classification of defect classes based on region feature vectors may be performed.

  Further, the inspection / classification apparatus 1 may be used for classifying cells in a predetermined liquid such as blood or culture liquid. In this case, a liquid containing cells is imaged to obtain a captured image (cell image). For example, in order to remove dead cells from cells in culture (for example, cancer cells), when a dead cell area is detected as a target area from a captured image showing the cells, the captured image is a dead cell area. In the dead cell image, the region of dead cells is specified by the user, and a binary dead cell region image is generated. The pixel feature quantity calculation unit 311 calculates a plurality of pixel feature quantities for each pixel of the dead cell image, and performs learning using the dead cell area image and the pixel feature quantity vector of each pixel to perform pixel classification. Vessel 312 is constructed. Then, when a captured image whose presence or absence and position of a dead cell region is unknown is input as a target image, whether or not the pixel is included in the dead cell region based on the pixel feature quantity vector of each pixel of the target image The area of dead cells in the target image is detected. The target region to be detected may be, for example, a region of non-target cells generated in ips cells in culture. Of course, classification of the target area based on the area feature vector may be performed.

  As described above, the inspection / classification apparatus 1 can easily detect a target area used for a predetermined purpose from target images obtained by imaging various types of objects. In the inspection / classification apparatus 1, the function as an area detection apparatus for detecting a target area may be used for various applications, and may be used separately from the imaging unit 2. In the inspection / classification apparatus 1, an image captured by a laser beam, an electron beam, an X-ray or the like may be inspected and classified.

  The configurations in the above embodiment and each modification may be combined as appropriate as long as no contradiction arises.

1 inspection / classification apparatus 6 window 9 substrate 61 mask pattern 70 target image 71 defect image 72 defect area image 73 detection area image 74 processed detection area image 311 pixel feature quantity calculation section 312 pixel classifier 321 area feature quantity calculation section 322 area Classifier 711, 711a Pixels S11 to S15, S21 to S26 Steps

Claims (10)

An area detection apparatus for detecting a target area used for a predetermined purpose from a target image, the area detection apparatus comprising:
A plurality of mask patterns different from one another are prepared, and each mask pattern smaller in size than the window is arranged at a plurality of positions in a window centered on each pixel of the target image, and the mask patterns are followed. A pixel feature amount calculation unit configured to obtain a plurality of pixel feature amounts respectively corresponding to the plurality of mask patterns by performing an operation;
A pixel classifier that determines whether each pixel is included in a target area based on the plurality of pixel feature amounts;
An area feature quantity calculation unit which calculates an area feature quantity for the target area detected by the pixel classifier;
An area classifier that classifies the target area based on the area feature amount;
Equipped with
A sum of pixel operation values of all the pixels included in the target area, wherein the region feature quantity calculation unit uses a value obtained by calculation according to a mask pattern for each pixel included in the target region as a pixel operation value. a region detection apparatus characterized that you calculated as the region feature. The area detection apparatus according to claim 1, wherein
An area detection apparatus characterized in that the plurality of mask patterns include mask patterns for high-order local autocorrelation features. The area detection device according to claim 1 or 2 , wherein
The area feature quantity calculation unit performs preprocessing for removing a micro area on an image that distinguishes the target area from other areas, and performs the pre-processing on the target area indicated by the pre-processed image. An area detection apparatus characterized by calculating a feature quantity. The area detection apparatus according to any one of claims 1 to 3 , wherein
An area detection apparatus characterized in that the target image is an image obtained by imaging an object, and the target area is an area indicating a defect on the object. The area detection apparatus according to claim 4 , wherein
The pixel classification is performed by performing learning using a defect area image acquired based on a difference image between a defect image indicating a defect and a predetermined reference image, and a plurality of pixel feature amounts for each pixel of the defect image. An area detection device characterized in that the device is constructed. A region detection method for detecting a target region to be used for a predetermined purpose from a target image, comprising:
a) A plurality of mask patterns different from each other are prepared, and in a window centered on each pixel of the target image, each mask pattern smaller in size than the window is arranged at a plurality of positions to make each mask pattern Acquiring a plurality of pixel feature amounts respectively corresponding to the plurality of mask patterns by performing the operation according to the plurality of pixel patterns;
b) determining whether each pixel is included in a target area based on the plurality of pixel feature quantities;
c) calculating an area feature amount for the target area detected in the step b);
d) classifying the target area based on the area feature amount;
Equipped with
In the step c), the sum of pixel operation values of all the pixels included in the target region is a value obtained by performing an operation according to a mask pattern for each pixel included in the target region as a pixel operation value, region detection method according to claim Rukoto calculated as the region feature. The area detection method according to claim 6 , wherein
A region detection method characterized in that the plurality of mask patterns include mask patterns for high-order local autocorrelation feature quantities. The area detection method according to claim 6 or 7 , wherein
The step c)
Pre-processing to remove a micro area to an image in which the target area is shown separately from other areas;
Calculating the region feature amount for the target region indicated by the preprocessed image;
A region detection method comprising: The area detection method according to any one of claims 6 to 8 , wherein
The area detection method, wherein the target image is an image obtained by imaging an object, and the target area is an area indicating a defect on the object. The area detection method according to claim 9 ,
The above b) is performed by learning using a defect area image acquired based on a difference image between a defect image indicating a defect and a predetermined reference image, and a plurality of pixel feature amounts for each pixel of the defect image. An area detection method characterized in that a pixel classifier used in the process is constructed.