TWI874236B - Classification method of defects in picture - Google Patents

Abstract

A classification method of defects in picture including the following steps is provided. Performing image alignment on an inspection image and a normal image, then subtracting the inspection image and the normal image, and take an absolute value of a brightness difference to obtain a difference image. Performing a high-sensitivity detection on the difference image to confirm whether there are defects in the inspection image. If the defects are detected in the high-sensitivity detection, the inspection image is inspected using a first classification method. If the defects are a first type or there is no classification result, an image line-width detection is performed on the normal image. If a line width of the normal image is larger than a first preset pixel size, a low-sensitivity detection is performed on the difference image to confirm whether there are the defects in the inspection image. If the line width of the normal image is less than or equal to the first preset pixel size and the defects have no classification result, the inspection image is inspected by a second classification method.

Description

Classification of defects in images

The present invention relates to a classification method, and in particular to a classification method for defects in an image.

During the chip manufacturing process, photolithography technology is used to copy the pattern on the mask onto the semiconductor wafer. Therefore, the quality of the mask affects the yield of chip manufacturing. In order to find defects that may affect the product, the mask is generally first put into the mask pattern inspection machine for inspection.

However, most of the inspection results of existing image inspection machines are false alarms (that is, they are not actually defects) and cannot classify defects. Therefore, manual secondary screening is required to eliminate false alarm results and classify defects one by one. Furthermore, directly using machine learning methods such as CNN/YOLO may miss high-risk defects and occasionally misclassify simple defect types.

The present invention provides a method for classifying defects in images, which can automatically classify inspection images with defects.

An embodiment of the present invention provides a method for classifying defects in an image, which includes the following steps. Perform image alignment on a test image and a normal image, then subtract the test image from the normal image and take the absolute value of the brightness difference to obtain a difference image. Perform high-sensitivity detection on the difference image to confirm whether there is a defect in the test image. When a defect is detected in the high-sensitivity detection, perform a first classification method detection on the test image. When the defect is of the first type or there is no classification result, perform image line width detection on the normal image. When the line width of the normal image is greater than a first preset pixel size, perform low-sensitivity detection on the difference image to confirm whether there is a defect in the test image. When the line width of the normal image is less than or equal to the first preset pixel size and the defect is a non-classified result, perform a second classification method detection on the test image.

Based on the above, in a method for classifying defects in an image of an embodiment of the present invention, high sensitivity detection or low sensitivity detection is used to confirm whether a defect exists in the test image, and then the test image is subjected to the first classification method detection or the second classification method detection. Moreover, the classification method further performs image line width detection to determine which classification method detection to perform. Therefore, the classification method of the embodiment of the present invention uses image processing combined with machine learning techniques to design the first and second classification methods to automatically classify defects in the image. In addition to having the advantages of high accuracy, low risk, and the ability to adjust the detection sensitivity according to the image line width, it can also reduce the burden of manpower and improve classification consistency.

FIG. 1A to FIG. 1C are schematic diagrams of defects of the first type and the first category to the third category of the first type, respectively. FIG. 2A to FIG. 2C are schematic diagrams of defects of the second type and the first category to the third category of the second type, respectively. FIG. 3A to FIG. 3B are schematic diagrams of defects of the third type and the first category to the second category of the third type, respectively. Please refer to FIG. 1A to FIG. 3B first. An embodiment of the present invention provides a method for classifying defects in a picture. The picture is, for example, an image of a pattern on a mask. However, the present invention is not limited thereto, and the method for classifying defects in a picture of the embodiment of the present invention can also be applied to other pictures, such as circuit patterns.

Taking the mask pattern as an example, the white area is a light-transmitting area, and its material is usually quartz (Qz). The black area is an opaque area, and its material is usually chromium (Cr). Among them, defects can be roughly classified into three types. The first type of defect is located at the junction of the black area and the white area in the picture. For example, Figure 1A shows that defect 1A is a black protrusion, Figure 1B shows that defect 1B is a white protrusion, and Figure 1C shows that defect 1C is an uneven junction between the black area and the white area (i.e., burrs). The second type of defect is located in the white area or the black area. For example, Figure 2A shows that defect 2A is a black dot in the white area, Figure 2B shows that defect 2B is a white dot in the black area, and Figure 2C shows that defect 2C is a black dot in the black area. The third type of defect is other types of defects. For example, FIG. 3A shows that defect 3A connects two black areas, and FIG. 3B shows that defect 3B may be a break in the black auxiliary line.

FIG4 is a flow chart of a method for classifying defects in an image according to an embodiment of the present invention. Referring to FIG4 , in this embodiment, the method for classifying defects in an image includes the following steps. Step S100, aligning the test image and the normal image, subtracting the test image from the normal image and taking the absolute value of the brightness difference to obtain a difference image. Step S200, performing high-sensitivity detection on the difference image to confirm whether there is a defect in the test image. Step S300, when a defect is detected in the high-sensitivity detection, performing the first classification method detection on the test image. Step S400, when the defect is of the first type or there is no classification result, performing image line width detection on the normal image. In step S500, when the line width of the normal image is greater than the first preset pixel size, a low-sensitivity detection is performed on the difference image to confirm whether the test image has a defect. In step S600, when the line width of the normal image is less than or equal to the first preset pixel size and the defect is a non-classified result, a second classification method detection is performed on the test image. The first preset pixel size is, for example, set to 10 pixels, but the present invention is not limited thereto.

In this embodiment, the high sensitivity detection in the above step S200 is to scan the difference map every 2 pixels × 2 pixels, and detect the position where the absolute value of the brightness difference is greater than or equal to the first threshold value. The image is stored in a grayscale of 0 to 255, for example. The first threshold value can be set to 20, but the present invention is not limited thereto. Moreover, the condition that the absolute value of the brightness difference is greater than or equal to the first threshold value can be set to be satisfied by each pixel or the average value of the pixels.

FIG5 is a flow chart of a first classification method in a method for classifying defects in a picture according to an embodiment of the present invention. FIG6 is a schematic diagram of classifying defects into a first category of a second form using the first classification method in a method for classifying defects in a picture according to an embodiment of the present invention. Please refer to FIG5 and FIG6. In this embodiment, the above-mentioned step S300 includes the following steps. Step S310, cropping the detection image IP and the normal image NP according to the position P of the defect. The above-mentioned cropping of the detection image IP and the normal image NP is, for example, a minimum square centered at the position P of the defect, and then extending outward to a square range of two pixels in size. In steps S320 to S330, when the absolute value of the brightness difference between the cropped detection image IP' and the cropped normal image NP' at the edge of the image is less than the first threshold value, and the brightness of the position P of the defect corresponding to the position of the normal image is greater than the second threshold value, it is determined that the defect at the position P is the second type first category 2A, as shown in FIG2A or FIG6. The second threshold value can be set to 180, but the present invention is not limited thereto.

In this embodiment, the above step S300 further includes the following steps. Steps S320 to S332: when the absolute value of the brightness difference between the cropped test image and the cropped normal image at the edge of the image is less than the first threshold value, the brightness of the position of the defect corresponding to the position of the normal image is less than or equal to the second threshold value, and the average brightness of the cropped normal image is less than or equal to the average brightness of the cropped test image, it is determined that the defect at the position is the second type, the second category 2B, as shown in FIG2B. In addition, the absolute value of the brightness difference between the cropped test image and the cropped normal image at the edge of the image is less than the first threshold value, the brightness of the position of the defect corresponding to the position of the normal image is less than or equal to the second threshold value, and the average brightness of the cropped normal image is greater than the average brightness of the cropped test image, and the defect is judged to be the second type third category 2C at this position, as shown in Figure 2C.

FIG7 is a schematic diagram of classifying defects into the first type and the first form using the first classification method in the method for classifying defects in an image of an embodiment of the present invention. Please refer to FIG5 and FIG7. In this embodiment, the above step S300 further includes the following steps. In steps S320 to S350, when the absolute value of the brightness difference at the edge of the cropped detection image IP1' or the cropped normal image NP1' is greater than or equal to the first threshold value, and the number of defects is less than 10, the cropped normal image NP1' is image segmented to segment the cropped normal image NP1' into a black area BR and a white area WR. In steps S360 to S382, when the absolute value of the brightness difference between the black area BR and the white area WR is greater than the third threshold value, the black area BR and the white area WR are horizontally or vertically segmented, the sum of the number of black areas BR and white areas WR is 2, and the average brightness of the cropped normal image NP1 is greater than or equal to the average brightness of the cropped test image IP1, it is determined that the defect at the position P1 is the first type and the first category 1A, as shown in FIG. 1A or FIG. 7. The method for performing image segmentation in the above step S350 can be to perform image segmentation using a machine learning method, such as using a k-means clustering algorithm. The third threshold value can be set to 50, but the present invention is not limited thereto. Moreover, when the number of defects detected in the test image IP is greater than or equal to 10, it is very likely that the system has a false alarm due to the high sensitivity of the detection or the poor image quality. Therefore, in step S340, the number of defects is limited to less than 10, and when the number of defects is greater than or equal to 10, it is temporarily judged as no classification result. Then, the detection sensitivity can be adjusted (i.e., step S500), or manual interpretation can be performed (i.e., step S640 in Figure 11).

In addition, in steps S360 to S382, when the absolute value of the brightness difference between the black area and the white area is greater than the third threshold value, the black area and the white area are divided horizontally or vertically, the sum of the number of black areas and the number of white areas is 2, and the average brightness of the cropped normal image is less than the average brightness of the cropped test image, it is determined that the defect at the position is the first type second category 1B, as shown in FIG. 1B.

Please refer to FIG5 , in this embodiment, the above step S300 further includes the following steps. Steps S360 to S392, when the absolute value of the brightness difference between the black area and the white area is greater than the third threshold value, the black area and the white area are divided horizontally or vertically, the sum of the number of black areas and the number of white areas is 3, and the opposite sides of the normal image after cropping (the upper and lower sides or the left and right sides are considered to be the black area and the white area are divided horizontally or vertically, such as the left and right sides in FIG3A ) are both black areas, it is determined that the defect at the position is the third type first category 3A, as shown in FIG3A .

FIG8 is a schematic diagram of classifying defects into the third form second category using the first classification method in the classification method of defect cases in pictures of an embodiment of the present invention. Please refer to FIG5 and FIG8, steps S360 to S394, when the absolute value of the brightness difference between the black area BR and the white area WR is greater than the third threshold value, the black area BR and the white area WR are divided horizontally or vertically, the sum of the number of black areas BR and white areas WR is 3, the opposite sides of the cropped normal image NP2' (for example, the left and right sides in FIG8) are both white areas WR, and the average brightness of the cropped normal image NP2 is less than the average brightness of the cropped detection image IP2, it is judged that the defect at the position P2 is the third form second category 3B, as shown in FIG3B or FIG8.

In addition, in steps S360 to S394, when the absolute value of the brightness difference between the black area and the white area is greater than the third threshold value, the black area and the white area are divided horizontally or vertically, the sum of the number of black areas and the number of white areas is 3, the opposite sides of the cropped normal image are both white areas, and when the average brightness of the cropped normal image is greater than or equal to the average brightness of the cropped test image, it is determined that the defect at the position is the first type, the first category 1A, as shown in FIG. 1A.

FIG9 is a schematic diagram of an example of performing image line width detection on a normal image in a method for classifying defect cases in an image according to an embodiment of the present invention. FIG10 is a schematic diagram of another example of performing image line width detection on a normal image in a method for classifying defect cases in an image according to an embodiment of the present invention. Please refer to FIG4, FIG9 and FIG10. Generally speaking, the line width of an image can be defined as the shortest interval between black areas. When the defect is classified as the first type and the line width is large, the system is prone to false alarms. Therefore, when the defect is classified as the first type, it is better to perform line width detection and then decide whether to adjust the detection sensitivity, i.e., steps S400 to S500.

In this embodiment, the above step S400 includes the following steps. Binarize the normal images NP3 and NP4, as shown in the left to middle of Figure 9 or Figure 10. Expand the binary normal images NP3' and NP4' by the second preset pixel size in the range of the black area BR', as shown in the middle to the right of Figure 9 or Figure 10. When the number of connections in the black area BR'' and the number of connections in the white area WR'' of the binary normal images NP3'' and NP4'' after expansion remain unchanged compared to before expansion (the number of connections in the black area BR and the number of connections in the white area WR of the normal images NP3 and NP4), it is judged that the line width is greater than or equal to the first preset pixel size. When the number of connections in the black area BR'' and the number of connections in the white area WR'' of the binary normal images NP3'' and NP4'' after expansion are less than before expansion (the number of connections in the black area BR and the number of connections in the white area WR of the normal image) (for example, in FIG. 10, the number of connections in the black area BR'' remains 1, but the number of connections in the white area BR'' changes from 5 to 0), it is determined that the line width is less than the first preset pixel size. The second preset pixel size is, for example, set to 5 pixels, but the present invention is not limited thereto.

That is, in the defect classification method of an embodiment of the present invention, when the line width is large, the detection sensitivity can be reduced. On the contrary, high sensitivity detection can be maintained. Therefore, by dynamically adjusting the detection sensitivity, the classification accuracy can be improved under low risk conditions.

In this embodiment, the low sensitivity detection in the above step S500 is to scan the difference map every 3 pixels × 3 pixels, and detect the position where the absolute value of the brightness difference is greater than or equal to the fourth threshold. The fourth threshold must be greater than the first threshold. For example, the fourth threshold can be set to 25, but the present invention is not limited to this. Moreover, the above-mentioned condition that the absolute value of the brightness difference is greater than or equal to the fourth threshold can be set to be satisfied by each pixel or the average value of the pixels.

FIG11 is a flow chart of a second classification method in a method for classifying defect cases in a picture according to an embodiment of the present invention. Please refer to FIG4 and FIG11. In this embodiment, the above step S600 includes the following steps. Step S610, using an object detection method to perform defect position detection on a normal image and a detection image around the position of the defect. Steps S620 to S630, when the defect position detected by the object detection method overlaps with the position of the defect, the object detection method is used to classify the defect. Among them, the object detection method can be the YOLO (You Only Look Once) object detection shown in FIG11, but the present invention is not limited thereto.

FIG. 12A to FIG. 12C are schematic diagrams of different types of defects detected by the second classification method in the method for classifying defect cases in pictures of an embodiment of the present invention. Taking FIG. 12A as an example, the detection image IP5 has a defect at position P5. Although this defect should be classified as the second type first category 2A, the edge of the black area is not a straight line. Therefore, this defect is classified as a non-classified result in the first classification method, as shown in steps S300, S400 and S600 or steps S300, S400, S500, S700 and S600 in FIG. 4. Taking FIG. 12B as an example, the detection image IP6 has a defect at position P6. Although this defect should be classified as Type 1, Type 2, 1B, the black and white areas in the normal image NP6 are not divided vertically or horizontally (the black and white junction is not a straight line, and the junction is uneven and burred), so it is also classified as a non-classified result in the first classification method. Taking Figure 12C as an example, the test image IP7 has a defect at position P7. Although this defect should be classified as Type 3, Type 1, 3A, the black and white junctions on both sides are not straight lines, so it is also classified as a non-classified result in the first classification method.

That is, when the first classification method is classified as unclassified and only the location information of the defect is available, the object detection method can be used for classification to increase the success rate of the automatic classification of the system. Moreover, when the types of some native patterns in the normal image are similar to the defects, the characteristics of the object detection method can be used to eliminate some misjudgment results. If the object detection method has no predicted results at the location of the defect, it can be handed over to manual classification, as shown in step S640.

FIG. 13 is a flow chart of a second classification method in a method for classifying defect cases in an image according to another embodiment of the present invention. Referring to FIG. 4 and FIG. 13, in another embodiment, the above step S600 may include the following steps. Step S610, using the object detection method to perform defect position detection on the normal image and the detection image around the position of the defect. Step S620', when the defect position detected by the object detection method in the normal image and the defect position in the detection image have the same classification and the range overlaps, remove the detection result of the object detection method. When the range of the defect position in the normal image overlaps with the range of the defect position in the detection image, A∩B/A∪B > 0.3, where A is the range of the defect position in the normal image, and B is the range of the defect position in the detection image.

FIG14 is a schematic diagram of removing detection results using the second classification method in the classification method of defect cases in pictures of another embodiment of the present invention. Taking FIG14 as an example, both the detection image IP8 and the normal image NP8 can be detected by the object detection method as having possible defects at the defect positions DP8 and DP8'. In other words, the native pattern in the normal image NP8 is similar to the type of the defect, which may cause a misjudgment by the object detection method. At this time, the object detection method can be used to detect the position of the object, and the detection results of the same classification that overlap (that is, the above-mentioned judgment method of overlap rate > 0.3) in the detection image IP8 and the normal image NP8 can be removed, thereby reducing the number of misjudgments and improving the accuracy of the system's automatic classification.

Please refer to FIG. 4 again. In this embodiment, the defect classification method in the image further includes the following steps. Steps S500 and S700: when the defect is detected in the low-sensitivity detection, the detection image is tested again using the first classification method. Steps S700 and S600: when the detection image is tested again using the first classification method and the defect is not classified, the detection image is tested using the second classification method.

In this embodiment, the defect classification method in the image further includes the following steps: Steps S500 and S800, when the defect is not detected in the low-sensitivity detection, the defect is determined to be of the first type third category 1C at its location, as shown in FIG1C .

In summary, the defects in the test image can be judged according to the rules of brightness, number of divisions, shape, etc. Defects with obvious morphological characteristics can be directly classified using the rules of the embodiment of the present invention, thereby improving the classification accuracy. For example, in the second type (the defect is located in the large black area or the large white area), if the brightness difference at the edge of the cropped image is <20 (the first threshold), the defect can be classified as 2A/2B/2C. In the first type (the defect is located at the junction of black and white) or the third type (other situations), if the black and white areas in the normal image are divided vertically or horizontally, the defect can be classified as 1A/1B/3A/3B.

In summary, in one embodiment of the present invention, the classification method of defects in an image is to perform high-sensitivity detection on the difference image to confirm whether there is a defect in the test image; when a defect is detected in the high-sensitivity detection, the test image is detected by the first classification method; when the defect is of the first type or there is no classification result, the normal image is detected by line width of the image; when the line width of the normal image is greater than the first preset pixel size, the difference image is detected at a low sensitivity to confirm whether there is a defect in the test image; and when the line width of the normal image is less than or equal to the first preset pixel size and the defect is a non-classified result, the test image is detected by the second classification method. Therefore, the classification method of the embodiment of the present invention uses image processing combined with machine learning techniques to design the first and second classification methods to automatically classify defects in images. In addition to its advantages such as high accuracy, low risk, and the ability to adjust detection sensitivity according to the line width of the image, it can also reduce manpower burden and improve classification consistency.

1A, 1B, 1C, 2A, 2B, 2C, 3A, 3B: Defects

BR, BR’’: Black Area

DP8, DP8’: Defect location

IP, IP1, IP2, IP5, IP6, IP7, IP8: Test Diagram

IP’, IP1’, IP2’: cropped detection images

NP, NP1, NP2, NP3, NP4, NP8: Normal image

NP’, NP1’, NP2’: normal images after cropping

NP3’, NP4’: normal images after binary transformation

NP3’’, NP4’’: Normal image after binary expansion

P, P1, P2, P5, P6, P7: Position

S100, S200, S300, S310, S320, S330, S332, S340, S350, S360, S370, S380, S382, S390, S392, S394, S400, S500, S600, S610, S620, S620’, S630, S640, S700, S800: Step

WR, WR’’: White area

Figures 1A to 1C are schematic diagrams of defects of the first type and the first category to the first type and the third category, respectively. Figures 2A to 2C are schematic diagrams of defects of the second type and the first category to the second type and the third category, respectively. Figures 3A to 3B are schematic diagrams of defects of the third type and the first category to the third type and the second category, respectively. Figure 4 is a flow chart of a method for classifying defects in a picture according to an embodiment of the present invention. Figure 5 is a flow chart of a first classification method in a method for classifying defects in a picture according to an embodiment of the present invention. Figure 6 is a schematic diagram of classifying defects into the first category of the second form using the first classification method in a method for classifying defects in a picture according to an embodiment of the present invention. Figure 7 is a schematic diagram of classifying defects into the first category of the first form using the first classification method in a method for classifying defects in a picture according to an embodiment of the present invention. FIG8 is a schematic diagram of classifying defects into the second category of the third form using the first classification method in the method for classifying defects in pictures according to an embodiment of the present invention. FIG9 is a schematic diagram of an example of detecting the line width of a normal image in the method for classifying defects in pictures according to an embodiment of the present invention. FIG10 is a schematic diagram of another example of detecting the line width of a normal image in the method for classifying defects in pictures according to an embodiment of the present invention. FIG11 is a flow chart of the second classification method in the method for classifying defects in pictures according to an embodiment of the present invention. FIG12A to FIG12C are schematic diagrams of different types of defects detected by the second classification method in the method for classifying defects in pictures according to an embodiment of the present invention. FIG. 13 is a flow chart of a second classification method in a method for classifying defects in a picture according to another embodiment of the present invention. FIG. 14 is a schematic diagram of removing detection results using the second classification method in a method for classifying defects in a picture according to another embodiment of the present invention.

S100, S200, S300, S400, S500, S600, S700, S800: Steps

Claims (11)

A method for classifying defects in an image is implemented by a processor, and includes: performing image alignment on a test image and a normal image, then subtracting the test image from the normal image and taking the absolute value of the brightness difference to obtain a difference image; performing high-sensitivity detection on the difference image to confirm whether the test image has a defect; when a defect is detected in the high-sensitivity detection, performing a first classification method detection on the test image, wherein the high-sensitivity detection is to scan the difference image every 2 pixels×2 pixels by the processor, and detect the position where the absolute value of the brightness difference is greater than or equal to a first threshold value; when the defect passes through the first classification method, When the detection method is the first type or no classification result, the normal image is subjected to image line width detection, wherein the line width is defined as the shortest interval of the normal image in the black area; when the line width of the normal image is greater than the first preset pixel size, the difference image is subjected to low sensitivity detection to confirm whether the detection image has the defect, wherein the low sensitivity detection is performed by the processor to scan the difference image every 3 pixels × 3 pixels and detect the position where the absolute value of the brightness difference is greater than or equal to the fourth threshold value; and when the line width of the normal image is less than or equal to the first preset pixel size and the defect is a no classification result, the detection image is subjected to the second classification method detection. The defect classification method in the image as described in claim 1, wherein the step of performing the first classification method detection on the detection image includes the following steps: Crop the detection image and the normal image according to the position of the defect; when the absolute value of the brightness difference between the cropped detection image and the cropped normal image at the edge of the image is less than the first threshold value, and the brightness of the position of the defect at the position of the normal image is greater than the second threshold value, it is judged that the defect at the position is the second type first category. The method for classifying defects in an image as described in claim 2, wherein the step of performing the first classification method on the test image further includes the following steps: the absolute value of the brightness difference between the cropped test image and the cropped normal image at the edge of the image is less than the first threshold value, the brightness at the position of the defect corresponding to the position of the normal image is less than or equal to the second threshold value, and the average brightness of the cropped normal image is less than or equal to the average brightness of the cropped test image. The average brightness of the defect is judged to be the second type and the second category at the position; and the absolute value of the brightness difference between the cropped test image and the cropped normal image at the edge of the image is less than the first threshold value, the brightness of the position of the defect corresponding to the position of the normal image is less than or equal to the second threshold value, and the average brightness of the cropped normal image is greater than the average brightness of the cropped test image, and the defect is judged to be the second type and the third category at the position. The defect classification method in the image as described in claim 1, wherein the step of performing the first classification method detection on the detection image further includes the following steps: cropping the detection image and the normal image according to the position of the defect; when the absolute value of the brightness difference of the cropped detection image or the cropped normal image at the edge of the image is greater than or equal to the first threshold value, and the number of defects is less than 10, performing image segmentation on the cropped normal image to divide the cropped normal image into the black area and the white area. The method for classifying defects in an image as described in claim 4, wherein the step of performing the first classification method on the test image further includes the following steps: the absolute value of the brightness difference between the black area and the white area is greater than a third threshold value, the black area and the white area are divided horizontally or vertically, the sum of the number of the black area and the white area is 2, and the average brightness of the cropped normal image is greater than or equal to the average brightness of the cropped test image. When the brightness is greater than or equal to 0, the defect at the position is judged to be of the first type and the first category; and when the absolute value of the brightness difference between the black area and the white area is greater than the third threshold value, the black area and the white area are divided horizontally or vertically, the sum of the number of the black area and the white area is 2, and the average brightness of the normal image after clipping is less than the average brightness of the test image after clipping, the defect at the position is judged to be of the first type and the second category. The method for classifying defects in an image as described in claim 4, wherein the step of performing the first classification method detection on the detection image further includes the following steps: when the absolute value of the brightness difference between the black area and the white area is greater than a third threshold value, the black area and the white area are divided horizontally or vertically, the sum of the number of the black area and the white area is 3, and both opposite sides of the normal image after cropping are the black area, it is judged that the defect at the position is the first type of the third type; when the absolute value of the brightness difference between the black area and the white area is greater than the third threshold value, the black area and the white area are divided horizontally or vertically, the sum of the number of the black area and the white area is 3, and the normal image after cropping is When the two opposite sides of the normal image are both white areas, and the average brightness of the normal image after clipping is less than the average brightness of the test image after clipping, the defect is judged to be the third type and the second category at the position; and when the absolute value of the brightness difference between the black area and the white area is greater than the third threshold value, the black area and the white area are divided horizontally or vertically, the sum of the number of the black area and the white area is 3, and the two opposite sides of the normal image after clipping are both white areas, and the average brightness of the normal image after clipping is greater than or equal to the average brightness of the test image after clipping, the defect is judged to be the first type and the first category at the position. The method for classifying defects in an image as described in claim 1, wherein the step of detecting the line width of the normal image includes the following steps: binarizing the normal image; expanding the binary normal image by a second preset pixel size in the range of the black area; when the number of connections in the black area and the number of connections in the white area of the binary normal image after expansion remain unchanged compared to before expansion, judging that the line width is greater than or equal to the first preset pixel size; and when the number of connections in the black area and the number of connections in the white area of the binary normal image after expansion decrease compared to before expansion, judging that the line width is less than the first preset pixel size. The method for classifying defects in images as described in claim 1, wherein the step of performing the second classification method detection on the detection image includes the following steps: using the object detection method to perform defect position detection on the normal image and the detection image around the position of the defect; and when the defect position detected by the object detection method overlaps with the position of the defect, classifying the defect using the object detection method. The defect classification method in the image as described in claim 1, wherein the step of performing the second classification method detection on the detection image includes the following steps: using the object detection method to perform defect position detection on the normal image and the detection image around the defect position; and when the defect position in the normal image detected by the object detection method is of the same classification as the defect position in the detection image and the range overlaps, removing the detection result of the object detection method, wherein when the range of the defect position in the normal image overlaps with the range of the defect position in the detection image, A∩B/A∪B>0.3, wherein A is the range of the defect position in the normal image, and B is the range of the defect position in the detection image. The method for classifying defects in an image as described in claim 1 further includes: when the defect is detected in the low-sensitivity detection, the detection image is tested again using the first classification method; and when the detection image is tested again using the first classification method and the defect is not classified, the detection image is tested using the second classification method. The method for classifying defects in an image as described in claim 1 further includes: when the defect is not detected in the low-sensitivity detection, judging that the defect is of the third category of the first type at its location.

Images