TW577995B - An image searching defect detector - Google Patents

Abstract

An image of an article to be inspected is divided into image portions, and a search engine makes a comparison with the image portion and a library of reference images. The reference images have predetermined labels that indicate whether each indicates a defect or no defect. The one of the reference images that most closely matches the image portion is determined, and the label associated with the reference image is taken as indicating whether the image portion corresponds to a location with a defect or no defect. Locations indicated as being defective are considered candidate defects and may subsequently be inspected in more detail.

Description

577995 A7 B7 V. Description of the invention (i) Cross-reference to related applications This patent application claims the benefit of US Provisional Patent Application No. 60 / 290,010 filed on May 11, 2001, which is incorporated by reference in its entirety. Incorporated herein. FIELD OF THE INVENTION The present invention relates to defect detection in inspecting electrical circuits. More specifically, the present invention relates to defect detection using an image classification and supervised learning system. BACKGROUND OF THE INVENTION (Electrical circuit inspection systems are well known and include, for example, INSPIRE-9060 ™ and SK-75 ™ automated optical inspection systems available from Yavne Orbotech Ltd., Israel. These electrical circuit inspection systems employ multiple Inspection channel: After obtaining a grayscale image or a color image of the electrical circuit under inspection, an enhanced resolution binary image of the electrical circuit is generated from the obtained image in the first inspection channel, and then analyzed to find out the electrical circuit Possible defects. Simultaneously, in the second inspection channel, the grayscale image is directly analyzed, usually detecting relatively small defects, such as pinholes, copper spots, and small short circuits. Grayscale image analysis is usually based on the reference image. A variety of grayscale characteristics are implemented. Characteristics include changes in grayscale values along various axes of the image part, such as grayscale values rising and / or decreasing. Usually it is not easy to see the defects detected by grayscale image analysis with the naked eye, which is not suitable People directly inspect the grayscale image. The reason for this difficulty is that the grayscale value may not change significantly to the naked eye, and Defects or • 4- This paper size applies Chinese National Standard (CNS) A4 (210 X 297 mm) 1 577995 A7 B7 V. Description of the invention (the number of possible non-defective gray-scale value changes and arrangements may be quite large. Invention SUMMARY The present invention seeks to provide an improved system and method for analyzing grayscale images to detect defects in objects such as printed circuit boards. According to a specific embodiment of the present invention, an image of a commodity under inspection is acquired and preferably is Subdivide or divide into a plurality of image parts. Each image part is supplied to an image search engine or image classifier for searching a database representing a plurality of labeled reference images to determine the corresponding damage of the image part. A defect or no defect. One way to accomplish this is to find the reference image or image that most closely resembles the image portion. At least one reference image of the reference images is marked to correspond to a defect, and the reference 5V The reference image to y of the image is marked as corresponding to a flawless pit. Provides a report of the reference image portion found most similar to a defective reference image * According to a specific embodiment of the present invention, the reference images "are represented by a plurality of typical features, each of which has a defined characteristic. For example, such features include gray levels along the -axis or above the image The characteristics of the intensity-increasing characteristics are the characteristics of these image parts, and the characteristics of the typical features of these parts are compared to represent the typical features of various reference image parts that are defective and non-defective. Based on the reference image of the image to determine-Image phase: 疋 Table 1 is defective or non-defective, for example, based on the analysis of selected features associated with each reference image. Pie ::: Pick: Fixed image part Typical features, and each feature refers to the value associated with the inspection image part. For example, it represents that the feature is relatively small relative to the gutter paper size -5- 577995 A7 B7 5. The value of the strength of the invention (3). The value of such a feature can be considered a defect detection value. The possible values of the feature set constitute a feature space. The feature space may be (and usually is) multi-dimensional. The values related to the features in the reference image collection are analyzed for defective and non-defective selected image representations, and are allocated between the spatial portion corresponding to the feature-related feature composition and the non-defect-related spatial composition. Feature space. The value of the feature related to the acquired image is plotted in the feature space, and a defect decision is made as to whether the position of the value related to an acquired image is part of a feature corresponding to a defect or to a non-defective feature. function. These image parts include, for example, a 5 X 5 pixel array. Individually estimated that each image of the product under test usually has between 10 and 5 < 9 images. According to a specific embodiment of the present invention, the image search engine or classifier is used as an initial defect filter, which is operated to select candidate image portions corresponding to possible defective parts in the inspected product, and thereby reduce the use for downstream images. Data processed. The portion of the image that is reported as corresponding to the defect can be supplied to, for example, a further processor that supplies a different (and often more accurate) set of defect detection algorithms. If necessary, obtain additional images corresponding to each image portion of a candidate defect (the resolution is usually higher than those of the image portions) and surrounding images, and analyze the additional images to determine whether the candidate defect It is indeed a defect or false alarm. According to a specific embodiment of the present invention, the image search engine is constructed and its sensitivity 纟 k is calibrated. For example, about 100 to 1000 image parts are initially reported as candidate defects. After further processing, they still report only There are about one mind image part as candidate defects. In addition, the image search engine is constructed and A4 size (210) 297 mm)

Binding

577995 A7 B7 V. Description of the invention (4) The sensitivity has been calibrated so as to evaluate the image portion of about 106 to 101 and find no defects. Only one image portion actually corresponds to the defect, but the error return corresponds to no defect. According to a specific embodiment of the present invention, it is preferable to select a labeled reference image by a supervised learning system, and an operator in the supervised learning system selects typical defective and non-defective image portions marked as defective or non-defective. Pre-determined characteristics and parameters of typical features are automatically taken offline by a suitable computer and stored in a database. The feature space segmentation for defective image parts and non-defective image parts can be revised to respond to the image parts added to the database. Brief description of the drawings The invention will be more clearly understood by reading the following detailed description and referring to the accompanying drawings, wherein: FIG. 1 shows a simplified functional block of a system for detecting defects in an image according to a specific embodiment of the invention Figure 2, Figure 2 shows a South-resolution black and white photo of a portion of an electrical circuit tested using the system shown in Figure 1, and Figure 3 shows a portion of the electrical circuit shown in Figure 2 obtained according to a specific embodiment of the present invention. Low-resolution digital images; Figures 4A and 4B show simplified diagrams of the operation of an image search engine used in the system shown in Figure 1 using the image portion used in Figure 3 and the system shown in Figure 1 according to a specific embodiment of the present invention; Figures 4C and 4D display According to another specific embodiment of the present invention, the image part used in FIG. 3 is used, and the image search used in the system shown in FIG. 1 is used. The paper size is applicable to the Chinese National Standard (CNS) A4 specification (210X 297 mm)

A simplified diagram of the engine's operation. In the second test, the display according to the embodiment of the present invention is used for \ drawing;:! The simplified functional block diagram of the system 10; please refer to Figure 2. The figure shows that the system shown in Figure 1 is not used, the selfless buckle and the electrical circuit of Shao Fen 12 ^ drunk listening black and white photos; Circle and touch Referring to Figure 3, Cai shows the digital image of "卩 m" shown in Figure 2 obtained according to the "instantaneous offense". Road-lower-resolution j According to a specific embodiment of the present invention, the system i 0 is operated to check the electrical circuit including the electrical circuit portion 12 shown in (example :), 2. Although the present invention has been described in terms of inspecting electrical circuit bodies, the systems and methods described herein can be applied to inspecting any properly manufactured commodity image, especially such as printed circuit boards, ball grid array substrates, multi-chip modules, and An electrical circuit such as a flat display. In addition, the term “electrical circuit” used in this article includes not only trimmed electrical circuits and printed circuit boards, but also electrical circuit layers that can be assembled together to form a complete electrical circuit. As shown in FIG. 1, the image generator 4 generates an image 1 6 (FIG. 3) of an inspected product, for example, an image of an electrical circuit part 2 (FIG. 2). Systems for obtaining digital images (such as image 16) and using these images (including processed representatives of electrical circuits derived from these images) to verify electrical circuits are well known. Examples of such commercially available printed circuit board inspection systems include, for example, the INSPIRE-9060TM & SK-75TM system, which can be addressed to Orbtech Ltd. of Yavne, Israel. According to a specific embodiment of the present invention, the system 10 divides a series of pure image parts. The paper size applies the Chinese National Standard (CNS) A4 specification (210 X 297 mm).

Binding

/ 799 $ / 799 $

18 is supplied to the feature extractor 20. Each pure image portion 18 is defined by a queue such as a 5 x 5 pixel array obtained from image 16 although a larger or smaller array can be used. The pure image portion 18 may be provided for the selected area of the product under test, or the pure image portion 18 may be provided as each pixel in the image 16 adjacent to the surrounding pixels. According to a specific embodiment of the present invention, a typical image of an electrical circuit includes pixels between lxlO9 and 5xl09. Between approximately 1x109 and 5xΐ09: Image portion 18 (each of which is a 5x5 pixel array Composition) is supplied to the feature extractor 20. It is desired to build the system 10 and obtain and analyze all the image parts of the tested electrical circuit in about 30 to 60 seconds by an operating system online. The picker 20 is preferably provided as a dedicated hardware device that is operable to retrieve one or more predetermined characteristic features 22 belonging to the characteristics of an image portion from the image portion 18 and facilitates further determination of the predetermined characteristics. Format to represent these features, as is well known in image search and classification techniques. In the background of grayscale analysis of image parts used in electrical circuit inspection, typical features of such images may be (for example) along the image part An indication of changes in one or more horizontal, vertical, or diagonal axis grayscale values. According to a specific embodiment of the present invention, a value is assigned to each feature in an image or Feature combination. For example, the value represents the relative strength of the features in the image. The feature of the present invention is that the typical feature 22 of the image 18 is supplied to the search engine 2 4 (such as a search engine or scene image classifier) for From an image library of appropriately labeled image parts similar to the defective and non-defective parts of the inspected product

577995 A7

28 receives an input 26. Obviously, the search engine 24 is operated on the defect and non-defective database to distinguish the characteristic representation of the image, and determine whether the image is a defect or non-defective representation. According to a specific embodiment of the present invention, the marked image portion is marked to indicate whether the image portion is an indication of whether the electrical circuit is defective or non-defective. It will be used in offline processing: and process the marked image part to capture typical features. According to the present invention, the embodiment of the 'image library 28 contains typical features and various other comparison parameters with weights and characteristic values, in order to facilitate the determination of the input image portion to the image portion stored in the image library 28 during the inspection process— Similarity. Binding

It is obvious that 'the image portion 18 obtained during the inspection and the image library will not completely match any particular image. According to a specific embodiment of the present invention' the search engine 24 is operated to find the input image portion 18 in the image library 28 which is most similar to Part of the image. As described above, images can be stored in the image library 28 as a set of typical features, a set of parameters used to determine the similarity of features, and a weight for mental importance, which is assigned to the input image. The exact similarity between the features of each part and the corresponding features in each image in the image library. The use of such features, parameters, and weighting is well known in search image and defect classification techniques. For each image part, the image search engine 2 4 searches the image library 2 8 'and identifies the image part 1 8 which is most similar to the image part 2 8 ^ ° is related to the image part of the image library 2 8 The mark indicates whether the obtained image portion 18 is defective or non-defective. The image search engine preferably provides a report of the image portion that most closely resembles the image library image marked as a defect30. If needed ’, a weight is provided to indicate the similarity of the image part to the 18 images in the image library.

577995 A7

577995 A7 B7 V. Description of the Invention (9) Please note that various image search engines on the market can receive input images and search the image collection to identify the images in the collection that are most similar to the input image. For example, it is also known to classify images for defect types. These search engines and classifiers often use neural networks, support vector machines, and / or decision tree techniques. A commercially available general-purpose image search engine is available from Attrasoft (www. Attrasoft.com) in Port Savannah, Georgia, USA. More specialized systems have also been used in automated optical inspection systems to inspect integrated circuits, semiconductor wafers, and liquid crystal displays, so that once defects are identified by the upstream image search engine, defects are then classified by defect type. One way is to think of search engines as image comparators. According to a specific embodiment of the present invention, the image search engine is constructed and its sensitivity is calibrated. For example, 100 to 1000 image parts are initially reported as candidate defects. After further processing, only about one image part is reported as candidate defects. . In addition, the image search engine is constructed and its sensitivity is calibrated to evaluate image portions of about 106 to 107 and found no defects. Only one image portion actually corresponds to the defect, but the error report corresponds to no defect. For example, the method of performing the calibration is to adjust respective comparison parameters, weights, and reliability of the features assigned to the image portions in the image library 28. According to a specific embodiment of the present invention, the image search engine 24 is used as a data reducer, and the report 30 is supplied to a downstream defect detector or defect verifier, which is a downstream defect detector or defect verifier It has a high degree of accuracy in detecting actual defects between image portions 18 in the acquired images that are reported as defects by the search engine. Downstream Defect Verifier -12- This paper size applies to China National Standard (CNS) A4 (210X 297mm) binding

Line 577995 V. Invention Description (10) 31 will provide a group defect detection algorithm different from the image search engine 24. If necessary, the downstream defect detector obtains an additional image of a location that is located on the electrical circuit being inspected and has been marked by the image search engine 24 as the location of a possible defect. For example, the resolution of the additional image may be higher than the resolution of the image portion 18. According to a specific embodiment of the present invention, the additional image is analyzed to determine whether the candidate defect reported by the image search engine 24 based on the analysis image portion 18 is indeed a defect indication, rather than an error warning indication. If necessary, according to another specific embodiment of the present invention, the defect verifier 3i is an operator verifying whether the defect reported in the report 30 is indeed a defect or no defect. The configuration of such a system 10 can be verified by human detection immediately after the search engine 24 is classified. If necessary, perform a manual verification operation after the automated downstream detection verification, so that the defect verifier 31 includes at least two verification operations, at least one automatic and at least one manual. According to a specific embodiment of the present invention, the image library 28, which can be regarded as a database of image-related data, is collected offline, and as described above, the images contained in the image library 28 are processed for the purpose of inspecting the images. Efficiently use images from Image Library 28 online. Therefore, as shown in Figure 丨, the image library 28 includes a collection of labeled reference images 32, which are collected offline and marked as corresponding to the defective and non-defective parts of the inspected product. Each reference image in the set of reference images 32 is provided to the offline parameter builder 34, and the parameter builder 34 extracts values related to features from each image in the image library, for example, corresponding to the defect decision value, This is a characteristic of the image portion contained in the collection. Parameter builder ________ -13 · This paper size applies to China National Standard (CNS) A4 (210 X 297 mm)

34 Assign one or more parameters for measuring the similarity or difference between the reference image part and the obtained image part, and the weighting of the importance belonging to the feature. As described above, the 'parameter building material is operated to define and segment the feature space table π of the respective strengths of the features in the defective image portion and the non-defective image portion. The feature space may be (and usually is) multi-dimensional. Please note that according to the specific implementation of the present invention, when searching for the image library image that most closely matches the input ~ image portion 18, the features and values that have been parameterized and stored in the image library 28 correspond to available image search engines 24 features that are easy and efficient to use. In alternative configurations, features correspond to features that can be evaluated by the image search engine to assign the inspected image to a feature space to determine defects and non-defects from the feature space. For example, in a system for inspecting printed circuit boards for small defects such as pinholes, copper spots, and small short circuits, such specific features may include grayscale values of one or more selected pixels in an image, And the trend and change of the grayscale value along each axis of an image part. For example,

The iNSPIRE-9060TM & SK-75TM ΑΙΙ machine purchased by Yavne Orbotech Ltd. uses the traditional characteristics of a 5x5 image portion suitable for small defects. The characteristics of small defects are extensively discussed and illustrated in one or more of the following U.S. patents, which are incorporated herein by reference in their entirety. U.S. Patent 5,619,429 and U.S. Patent 5,586,058, but please note that image classification is used in The actual composition and parameters of the features are not part of the present invention 'and please note that' any applicable feature can be captured, stored and applied to the classification of the image portion 18. In addition, any suitable size can be adopted. -14- This paper size is applicable to China National Standard (CNS) A4 specification (210X297 mm) 577995

Image search method. Reference is now made to Fig. 2, which shows an enlarged photograph of a portion of a typical electrical circuit that can be inspected with reference to the system shown and illustrated in Fig. 丨. The circuit portion 12 shown in FIG. 2 includes two conductors 40 and 42. A slight short circuit can be seen between the conductors and ^ shown in the figure, and a prominent protrusion can be seen along the left side of the conductor 40. Now refer to FIG. 3, which shows a digital image of the circuit portion 12, such as a digital image commonly available in commercially available electrical circuit inspection systems. Obviously, the resolution of the digital image 16 is extremely lower than that of the enlarged photo 丨 2, and it is difficult to identify the slight short circuit or protrusion by directly evaluating the digital image 6 with the naked eye. As shown in FIG. 3, the characteristic of a slight short circuit or protrusion is that the grayscale values of several pixels are different from the expected grayscale values of the pixels along the conductor 40 and 42 side or the defect-free segment between the conductors 40 and 42. The difference can be seen by comparing the pixels in the digital image 16 corresponding to the adjacent parts along the conductors 40 and 42. Referring now to Figs. 4A and 4B, and Figs. 4 (: and 41), the figure shows a simplified diagram of the operation of the image search engine 24 employed in the system shown in Fig. 1. In FIGS. 4A to 4D, the operation of the image search engine 24 will be explained using the image portion shown in FIG. In Figs. 4A to 4D, the knowledge base 50 (such as the image base 28) of the images 52 to 64 is prepared in an offline learning mode and is stored as a set of parameters, such as the parameters described above with reference to Fig. 丨. Obviously, images 5 2 to 6 4 are rendered as pure images based on simplification. In addition, for the purpose of simplifying illustration, the images 52 to 64 are widely plotted in a two-dimensional image space in order to regard the knowledge base 50 as a feature space. However, please note that at -15- this paper size applies the Chinese National Standard (CNS) A4 specification (210 X 297 public love)

Binding

line

577995 A7 B7 V. Description of the invention (13) In the diagrams shown in FIGS. 4A to 4D, no attempt is made to organize the images 52 to 64 in a symmetrical manner according to any specific feature, or the multidimensional nature of the feature space is attempted. & In Figs. 4C and 4D, the partition area 66 shown in the figure divides the knowledge space into the feature space portion of the image corresponding to the defect and the feature space portion of the image corresponding to the defect. Each of the image parts 5 2 to 6 4 represents a part of the circuit shown in FIG. 2. The image portions 52 to 60 belong to a portion corresponding to the obtained image 16 which is defect-free in the electric circuit portion 12 (Fig. 2). The image parts u and 64 belong to a part of the obtained image corresponding to the defect in the electrical circuit part 12. Please note that images 52 to 64 are indications of possible defects, but do not include all possible image portions and defects that would appear in the image. This is because the image search engine 28 is operated to find the most closely matched defective or non-defective image portion of the obtained image portion, or to determine whether the obtained image portion is located in the image space associated with the defect representation or non-defective representation. . Obviously, therefore, the image search engine does not need an exact match to determine each image part obtained. In fact, the excessive image portion in the knowledge base is not conducive to the scene > high-efficiency functions like the search engine 24. As shown in FIG. 4A (which is equivalent to the _ mode operation), the obtained image portion 70 (for example, 'the image portion obtained by the image generator 14) is compared with each of the images 5 2 to 6 4❶ Indicates whether the image part 70 is similar to: mesh = image parts 52 to 64 'and indicates whether the image part 70 is most similar to which of the two image parts in the image library 50. As shown in the example shown in the example, the image portion 70 is most similar to the image portion M marked as a defect. Therefore, the paper size is divided into two specifications (21〇χ 2 ^ γ -16 * 577995 Α7 Β7 V. Description of the Invention (M) Report the scene > image portion 70 as a defect and further process it. As shown in Figure 4B (which is also equivalent to the first mode operation), the obtained image portion 72 (for example , Using the image portion obtained by the image generator 14) to compare each of the images 52 to 64. Indicate whether the image portion 72 is similar or dissimilar to the image portions 52 to 64, and indicate that the image portion 72 is the most similar. Which image part is similar to the knowledge base 50. As shown in the example of FIG. 4B, the image part 72 is determined to be most similar to the image part 54 marked as non-defective. Yu Xie reports the image part 72 as No defect, so no further processing is needed. As shown in Figures 4C and 4D (which is equivalent to the second mode operation), the knowledge base 50 is provided in the image space where the image is located as a given feature in the image. Composition function of relative intensity. Image space is divided into 66. points So that the image of the defective portion is located in the image space on the first side of the partitioned area 66, and the image of the non-defective portion is located in the image space on the other side of the partitioned area 66. According to a specific embodiment of the present invention, each image All are represented by various feature intensity composition, and the position of the partition 66 in the image space is determined analytically as a value function. As shown in Figure 4C, the image part 70 is plotted in the knowledge base " The position in the feature space is shown. Since the image part 70 is located in the image space of the image space related to the defect, the image part 70 is reported as corresponding to the defect as shown in FIG. 4D and the image part 72 The bit m plotted in the feature space of the knowledge base 50 is the shadow component 72, which is located with ▲ Defect = "the part of the image space ', so the image part 72 is reported as corresponding to -17-

Claims (1)

A method for detecting defects in electrical circuits by gels, including ... obtaining images corresponding to a part of the circuit under test;-comparing the material and reference materials, such references At least two reference images of the image represent a defect, and at least _ images of the reference images represent a defect-free; and based on the comparison, provide an indication of the similarity of the acquired image to the defect or absence of defects β 2. If a patent is applied ㈣ The method of item 'wherein the part of the larger image of the electrical circuit under test. ， 象 疋 又 3. The method according to the first patent application range, in which one or more typical features are captured for the obtained image. 4. The method according to item 2 of the scope of patent application, in which one or more typical features are extracted for each reference image, and the typical features are stored in an image database. 5. The method according to item 4 of the scope of patent application, wherein the feature extracted from one of the reference images is related to a value indicating the intensity belonging to such feature, and an image part is obtained by at least two A combination of the values of the features. 6. The method of claim 5, wherein the comparison includes plotting the position of an image in a feature space as a function of the combination of the values. 7. The method according to item 6 of the scope of patent application, wherein the feature space includes a portion corresponding to a combination of values of defect indications, and corresponding to a non-defective index. The combination of the values shown. 8. If the scope of the patent application is No. 1 Lingshi, +, ^, and Yiwanfa, the method further includes using a Shishi, 4 »., ^^ wanfa that is different from the comparison. At the location marked as an image of a defect, the batch / v performs a subsequent inspection of the inspected electrical circuit from above. 9. The method of claim 1 in the patent application scope, the method further comprising: 'for-corresponding to a position corresponding to the image marked as a defect of the comparison' to obtain an additional image; and using a different image from the comparison; Inspection method to inspect the additional image. 10. The method of claim 9 in which the resolution of the obtained image is lower than the resolution of the additional image. 11. The method according to item 丨 of the patent application, wherein the resolution of the obtained image is insufficient to detect at least some defects with the naked eye. “12. If the method of applying for the first item of the patented red enclosure is applied, Langzhong will perform the comparison for at least 109 images during the detection of defects in the electrical circuit. 13. · An electrical circuit inspection system, comprising: an image generator for generating an acquired image, the acquired image corresponding to a portion of an electrical circuit under inspection; an image comparator, which is operated to compare the office Obtain an image and at least two reference images, wherein at least one reference image of the at least two reference images represents a defect in an image of an electrical circuit, and at least one reference image of the at least two reference images represents an electrical One defect in one of the images of the circuit; and a detection indicator to indicate that the obtained image is for the electrical circuit -20- This paper size applies the Chinese National Standard (CNS) A4 specification (210X297 mm) 577995 A8 B8 C8 ~ D8 6. The similarity of one defect or one defect in the scope of patent application. 14. A method of manufacturing an electrical circuit, comprising: supplying electrical conductors on a non-conductive substrate to form part of a τ electrical circuit; and inspecting a portion of the electrical circuit under inspection, including: generating an acquired image, the laboratory The obtained image corresponds to a part of the electrical circuit under test; comparing the obtained image with at least two reference images, wherein at least one reference image of the at least two reference images represents a defect in an electrical circuit, and the At least one reference image of the at least two reference images indicates a defect-free electrical circuit; and an indication is provided to indicate the similarity of the obtained image to a defect or a defect-free electrical circuit under test. 15. —A method for detecting defects in electrical circuits, including: a collection of images that form parts of the electrical circuits, the parts being classified at least as corresponding to one of a defect and a defect-free one; The images in the set are assigned defect decision values related to the characteristics representing the images, and the defect decision values are stored in a memory; an acquired image is generated, the acquired image corresponding to a A part; assigning a defect decision value to one of the features in the acquired image; comparing the stored defect decision values to an acquired image-21- This paper scale applies Chinese National Standard (CNS ) A4 specification (210 X 297 mm) 577995 A BCD Defect decision value related to the scope of patent application; and based on the analysis' provide an indication of the similarity of the obtained image to defects or no defects. 16. The method according to item 15 of the scope of patent application, wherein the composition-image collection includes selectively adding additional images that are classified as corresponding to a defect into the collection. ^ Π. The method according to item 16 of the patent application scope, wherein the composition_image collection includes selectively adding to the collection additional images that are classified as corresponding to a defect-free. 18. The method according to item 15 of the scope of patent application, wherein the composition_image collection includes selectively adding additional images classified as corresponding to a defect-free to the collection. 19. The method of claim 15 in the scope of patent application, wherein the image of a part of the electrical circuit is a larger image portion of an electrical circuit. 20. The method of claim 15 of the scope of patent application, the method further comprising using-a test method different from the comparison,-at a position corresponding to the image marked as equivalent to a missing Fa, performing the test of the electrical circuit Follow-up inspection. 21. The method of claim 15 in the scope of patent application, the method further comprising: obtaining an additional image for a position corresponding to the acquired image marked as a defect by the analysis; and using a comparison different from the comparison Inspection method to inspect the additional image. 22. The method of claim 21, wherein the resolution of the obtained image is lower than the resolution of the additional image. This paper size is applicable to China Standard (CNS) A4 (G10 × 297mm) -22 · 577995 A8 23. The method according to item " of the scope of patent application, wherein the assignment of the defect decision values is to assign the defect decision values in the feature space. 24. ^ A method for declaring patent No. 23, the method further comprising allocating a pan-feature workshop to build a part of the feature space belonging to the defect representation, and a part of the feature space belonging to the non-defect representation. . 25. If the method according to item 23 of the scope of patent application, the method further comprises: selectively adding additional images classified as corresponding to a defect to the set; and adjusting the feature space portion such as children to respond Selectively added image. 26. If the method of the scope of patent application 24, wherein the analysis includes determining the position of the defect decision values related to the _ obtained image in the feature space ® 27 · If the method of scope 15 of the patent application, The defect decision values represent the relative strength of a corresponding feature in an image. -23 · This paper size applies to China National Standard (CNS) A4 (210 X 297 mm)