Image Processing Method, Non-Transitory Computer Readable Medium, and Image Processing System

Abstract

An image processing method, a non-transitory computer-readable medium, and an image processing system. The image processing method includes the following operations: obtaining a defect image of a wafer; processing the defect image and generating a reconstructed image; and when the reconstructed image includes at least one object pattern, outputting the reconstructed image. The object pattern corresponds to a part of the wafer.

Description

Technical Field

[0001] The present disclosure relates to an image processing method and an image processing system, and more particularly to a method and a system for training an image processing program to generate a target image. Background Art

[0002] An optical imaging device and an electron microscope cooperate and are applied to inspect defects of a semiconductor wafer. Since the wafer image captured by the optical imaging device is not clear enough, and since it takes a lot of manpower and time to further inspect this wafer image with the electron microscope, the efficiency of inspecting the defects of the wafer is affected. Summary of the Invention

[0003] An embodiment of the present disclosure is about an image processing method. The image processing method includes the following operations: obtaining a defect image of a wafer; processing the defect image and generating a reconstructed image; and when the reconstructed image includes at least one object pattern, outputting the reconstructed image. The object pattern corresponds to a part of the wafer.

[0004] An embodiment of the present disclosure is about a non-transitory computer-readable medium. The non-transitory computer-readable medium includes executable instructions for implementing an image processing method by a processor. The image processing method includes the following operations: capturing a defect image, where the defect image includes a defect pattern corresponding to a wafer; and generating a reconstructed image according to the defect image and a plurality of first reference images, where the reconstructed image includes at least one object pattern corresponding to the wafer, and the first reference images include a plurality of patterns corresponding to the wafer. The object pattern and the defect pattern correspond to the same part of the wafer, and the object pattern is different from the defect pattern.

[0005] An embodiment of the present disclosure is about an image processing system. The image processing system includes a memory and a processor. The memory is used to store an image processing program. The processor is coupled to the memory. The processor is used to access the image processing program in the memory to perform the following steps: training the image processing program using a plurality of reference images; and processing a defect image through the trained image processing program to generate and output a reconstructed image. The reconstructed image includes at least one object pattern corresponding to a part of the wafer, and the reference images include a plurality of reference object patterns corresponding to a plurality of parts of the wafer. Brief Description of the Drawings

[0006] Aspects of some embodiments of the present disclosure will be best understood when read in conjunction with the accompanying drawings. Note that, in accordance with standard practice in the industry, various features are not drawn to scale. In fact, for clarity of discussion, the dimensions of various features may be arbitrarily increased or decreased.

[0007] According to some embodiments of the present disclosure,Figure 1 It is a schematic diagram of an image processing system;

[0008] According to some embodiments of the present disclosure, Figure 2A is a schematic diagram of a defective image corresponding to Figure 1 the defective image;

[0009] According to some embodiments of the present disclosure, Figure 2B is a schematic diagram of a reconstructed image corresponding to Figure 1 the reconstructed image;

[0010] According to some embodiments of the present disclosure, Figure 2C is a schematic diagram of a first reference image corresponding to Figure 1 the first reference image;

[0011] According to some embodiments of the present disclosure, Figure 2D is a schematic diagram of a second reference image corresponding to Figure 1 the second reference image;

[0012] According to some embodiments of the present disclosure, Figure 3 is a schematic diagram of an image processing program corresponding to Figure 1 the image processing program;

[0013] According to some embodiments of the present disclosure, Figure 4 is a flowchart of an image processing method;

[0014] According to some embodiments of the present disclosure, Figures 5A to 5B is a flowchart of a method for training an image processing program corresponding to Figure 4 the image processing method;

[0015] According to some embodiments of the present disclosure, Figure 6 is a flowchart of an image processing method;

[0016] According to some embodiments of the present disclosure, Figure 7 is a schematic diagram of an operation for classifying defective images corresponding to Figure 6 the image processing method;

[0017] According to some embodiments of the present disclosure, Figures 8A to 8C is a schematic diagram of a defective image corresponding to Figure 7 the classification result;

[0018]

Symbol Description

[0019] 100: Image processing system

[0020] 110: Image processing device

[0021] 11a: Memory

[0022] 11b: Processor

[0023] 111: Image processing program

[0024] 112: Reconstructed image

[0025] 120: Electronic device

[0026] 121: Reference image

[0027] 130: Electronic device

[0028] 131: Reference image

[0029] 132: Defect image

[0030] 200A: Defect image

[0031] 200B: Reconstructed image

[0032] 200C: Reference image

[0033] 200D: Reference image

[0034] 210: Defect pattern

[0035] 211: Background pattern

[0036] 220: Object pattern

[0037] 221: Background pattern

[0038] 230: Reference object pattern

[0039] 231: Reference background pattern

[0040] 240: Reference defect pattern

[0041] 241: Reference background pattern

[0042] 311: Image generation model

[0043] 312: Image discrimination model

[0044] 400: Method

[0045] S410, S420, S430, S440: Operations

[0046] 500A, 500B: Methods

[0047] S511, S512, S513, S514, S515, S516, S517, S518: Operations

[0048] S521, S522, S523, S524, S525: Operations

[0049] 600: Method

[0050] S610, S620, S630, S640, S650: Operations

[0051] 711, 712, 721, 722, 723, 724, 731, 732, 733: Blocks

[0052] 800A, 800B, 800C: Defect Images Detailed Implementation Manner

[0053] The following disclosure provides many different embodiments or examples for implementing different features of the provided subject matter. Specific examples of components and configurations will be described below to simplify some embodiments of the present disclosure. Of course, these are only examples and are not intended to be restrictive. For example, in the following description, the formation of a first feature above or on a second feature may include embodiments where the first feature and the second feature are formed in direct contact, and may also include embodiments where additional features may be formed between the first feature and the second feature such that the first feature and the second feature are not in direct contact. Additionally, some embodiments of the present disclosure may repeat reference numerals and / or letters in various examples. This repetition is for simplicity and clarity purposes and does not itself prescribe the relationship between the various embodiments and / or configurations discussed.

[0054] The terms used in this specification generally have the ordinary meanings of the terms in the art and in the specific context where each term is used. The use of examples in this specification (including examples of any terms discussed herein) is merely illustrative and in no way limits the scope and meaning of the present disclosure or any of the exemplified terms. Similarly, the present disclosure is not limited to the various embodiments given in this specification.

[0055] Although terms such as "first", "second", etc. may be used herein to describe various elements, these elements should not be limited by these terms. These terms are used to distinguish one element from another. For example, without departing from the scope of the embodiments, a first element may be referred to as a second element, and similarly, a second element may be referred to as a first element. As used herein, the term "and / or" includes any and all combinations of one or more of the associated listed items.

[0056] In this document, the term "coupled" may also be referred to as "electrically coupled", and the term "connected" may be referred to as "electrically connected". "Coupled" and "connected" may also be used to indicate that two or more elements cooperate or interact with each other.

[0057] In addition, for convenience in the description of the relationship between one element or feature illustrated in the figures and another element or feature, spatial relative terms may be used herein, such as "under", "below", "lower", "above", "upper", and the like. Spatial relative terms are intended to cover different orientations of an element during use or operation in addition to the orientation depicted in the figures. The structure may be oriented otherwise (e.g., rotated 90 degrees or in other orientations), and the spatial descriptors used herein may be interpreted accordingly.

[0058] As used herein, "about", "approximately", "nearly", or "substantially" generally refers to any approximation of a given value or range, where the approximation varies depending on the various fields to which it pertains, and the scope of the approximation should conform to the broadest interpretation understood by one of ordinary skill in the art to which the approximation pertains, so as to cover all such modifications and similar constructions. In some embodiments, the approximation generally means within 20%, preferably within 10%, and more preferably within 5% of a given value or range. The numerical quantities given herein are approximate, which means that the terms "about", "approximately", "nearly", or "substantially" may be inferred without being explicitly stated, or imply other approximations.

[0059] The operation of a semiconductor manufacturing process includes wafer fabrication and wafer defect review. In some embodiments, the defect review is used to image a wafer and generate an image of the wafer. In some embodiments, the defect review is further used to determine whether the wafer is defective based on the image. Some embodiments of the present disclosure provide systems and methods that are applied to defect review.

[0060] Now refer to Figure 1 . According to some embodiments of the present disclosure, Figure 1 is a schematic diagram of an image processing system 100. The image processing system 100 includes an image processing device 110, a first electronic device 120, and a second electronic device 130. The image processing device 110 is coupled to the first electronic device 120 and the second electronic device 130.

[0061] The image processing device 110 includes a memory 11a and a processor 11b. The memory 11a is coupled to the processor 11b. In some embodiments, the memory 11a is used to store an image processing program 111 and a reconstructed image 112. In some embodiments, the memory 11a is a non-transitory computer-readable medium and is used to store computer program code (i.e., executable instructions) (not shown in the figure). In various embodiments, the processor 11b is used to access the image processing program 111 and / or the computer program code (not shown in the figure) in the memory 11a to perform operations related to image processing. In various embodiments, the image processing device 110 is implemented by using a device with computing capabilities, such as a computer.

[0062] The first electronic device 120 is used to generate one or more first reference images 121. In some embodiments, the first electronic device 120 is used to photograph a wafer to generate corresponding one or more photos. In some embodiments, the one or more photos are used to train the image processing program 111, and the one or more photos are also referred to as the first reference images 121. In various embodiments, the first electronic device 120 is implemented by using a device that uses electron beam imaging. For example, the first electronic device 120 is implemented by a scanning electron microscope (SEM).

[0063] The second electronic device 130 is used to generate one or more second reference images 131 and one or more defect images 132. In some embodiments, the second electronic device 130 is used to photograph a wafer to generate corresponding one or more photos. In some embodiments, the one or more photos are used to train the image processing program 111, and the one or more photos are also referred to as the second reference images 131. In some other embodiments, the one or more photos are used to be analyzed or processed by the trained image processing program 111, and the one or more photos are also referred to as the defect images 132. In various embodiments, the second electronic device 130 is implemented by using a device that uses optical imaging. For example, the second electronic device 130 is implemented by an optical scanner with multiple wavelengths.

[0064] In some embodiments, the memory 11a is also used to store the defect images 132 captured by the second electronic device 130. In various embodiments, the processor 11b is used to access the image processing program 111 and the defect images 132 in the memory 11a to perform image processing operations. The image processing operations are described in more detail below with at least reference to Figures 3 to 4 be described in more detail.

[0065] In some embodiments, the memory 11a is further configured to store a first reference image 121 captured by the first electronic device 120 and a second reference image 131 captured by the second electronic device 130. In various embodiments, the processor 11b is configured to access the image processing program 111, the first reference image 121, and the second reference image 131 in the memory 11a to perform operations for training the image processing program 111. The operations of the training image processing program 111 are described in more detail below with reference to at least Figures 5A to 5B as follows.

[0066] For illustration purposes, some embodiments of the present disclosure provide Figure 1 the configuration of the image processing system 100, but not limited thereto. Any configurations for implementing Figure 1 the image processing system 100 are within the scope contemplated by some embodiments of the present disclosure. For example, in some embodiments, at least one of the first electronic device 120 or the second electronic device 130 is integrated with the image processing device 110. In various embodiments, the first electronic device 120 is coupled to the second electronic device 130.

[0067] Now refer to Figures 2A to 2B . According to some embodiments of the present disclosure, Figure 2A is a schematic diagram of a defective image 200A. In some embodiments, the defective image 200A is an example of the defective image 132 shown in Figure 1 . According to some embodiments of the present disclosure, Figure 2B is a schematic diagram of a reconstructed image 200B. In some embodiments, the reconstructed image 200B is an example of the reconstructed image 112 shown in Figure 1 .

[0068] As Figure 2A shown, in some embodiments, the defective image 200A is an optical image captured by a light beam and is a magnified partial image of one of the multiple dies of a wafer. In other words, referring to Figure 1 and Figure 2A , the image generated by capturing a part of the wafer by the second electronic device 130 (e.g., an optical scanner device) is the defective image 200A.

[0069] As Figure 2AAs shown, the defect image 200A includes a defect pattern 210 and a background pattern 211. In some embodiments, the defect pattern 210 corresponds to a part of the wafer, and the background pattern 211 corresponds to another part of the wafer. More specifically, the defect pattern 210 is a pattern in the image of the wafer and is a pattern of the defect of the wafer. The background pattern 211 is a pattern in the image of the wafer and is a pattern of the surface of the wafer. The defects of the wafer may include, for example, residues, scratches, or cracks. The surface of the wafer may include, for example, the surface or back surface of one of the plurality of metal layers, the surface or back surface of the active region and the electrodes, etc.

[0070] As Figure 2B shown, in some embodiments, the reconstructed image 200B is an image simulated by an electron beam capture, where this image is a magnified partial image of one of the plurality of die of the wafer. In other words, referring to Figure 1 and Figure 2B , the image generated by processing through the image processing program 111 is the reconstructed image 200B.

[0071] As Figure 2B shown, the reconstructed image 200B includes an object pattern 220 and a background pattern 221. In some embodiments, the object pattern 220 corresponds to a part of the wafer, and the object pattern 220 and the defect pattern 210 correspond to the same part of the wafer. In other words, the object pattern 220 and the defect pattern 210 are patterns of the same part (i.e., the defect) of the wafer. Since the defect image 200A and the reconstructed image 200B are images generated in different ways, the object pattern 220 is different from the defect pattern 210. Therefore, the defect image 200A does not have the object pattern 220, and the reconstructed image 200B does not have the defect pattern 210.

[0072] Similarly, in some embodiments, the background pattern 221 in the reconstructed image 200B is relative to the background pattern 211 in the defect image 200A. The background pattern 221 and the background pattern 211 correspond to the same part of the wafer (i.e., the wafer surface), but the background pattern 221 is different from the background pattern 211.

[0073] Now refer to Figures 2C to 2D . According to some embodiments of the present disclosure, Figure 2C is a schematic diagram of a first reference image 200C. In some embodiments, the first reference image 200C is an instance of the first reference image 121 as Figure 1 shown. According to some embodiments of the present disclosure, Figure 2D is a schematic diagram of a second reference image 200D. In some embodiments, the second reference image 200D is an instance of the second reference image 131 as Figure 1 shown.

[0074] As Figure 2CAs shown, in some embodiments, the first reference image 200C is an image associated with a wafer taken under a certain condition. This condition indicates that the first reference image 200C is an image captured by an electron beam and is a partial enlarged image of one of the multiple die of the wafer. In other words, referring Figure 1 and Figure 2C , the image generated by photographing a part of the wafer through the first electronic device 120 (for example, an SEM device, hereinafter referred to as the first electronic device 120) is the first reference image 200C. In some embodiments, the imaging method of the first reference image 200C is different from that of the reconstructed image 200B in Figure 2B . As described above, in some embodiments, the reconstructed image 200B is an image generated by processing the defect image 200A through the image processing program 111.

[0075] As Figure 2C shown, in some embodiments, the first reference image 200C includes a reference object pattern 230 and a reference background pattern 231. In some embodiments, the reference object pattern 230 corresponds to a part of the wafer, and the part of the wafer corresponding to the reference object pattern 230 is the same as that corresponding to the defect pattern 210. In some other embodiments, the parts of the wafer corresponding to the reference object pattern 230, the defect pattern 210, and the object pattern 220 are the same. In other words, the defect pattern 210, the object pattern 220, and the reference object pattern 230 are patterns of the same part (i.e., the defect) of the wafer. For example, referring Figures 1 to 2C , the defect pattern 210 corresponds to a defect of the wafer and is a pattern in the image generated by using the second electronic device 130 (for example, an optical scanner device, hereinafter referred to as the second electronic device 130). The reference object pattern 230 also corresponds to this defect and is a pattern in the image generated by using an SEM device. The object pattern 220 also corresponds to this defect of the wafer and is a pattern in the image simulated by using the image processing program 111.

[0076] Similarly, in some embodiments, the reference background pattern 231 in the first reference image 200C corresponds to the same part (i.e., the surface) of the wafer as at least one of the background pattern 211 in the defect image 200A or the background pattern 221 in the reconstructed image 200B.

[0077] The second reference image 200D is illustrated in Figure 2D, and the second reference image 200D includes a reference defect pattern 240 and a reference background pattern 241. In some embodiments, the second reference image 200D is an image associated with a wafer taken under a certain condition. This condition indicates that the second reference image 200D is an image captured by an optical scanner device and is a partial enlarged image of one of the multiple die of the wafer. In some embodiments, the second reference image 200D corresponds to Figure 2A the defect image 200A therein, and the similarities are not elaborated herein.

[0078] In some embodiments, the first reference image 200C and the second reference image 200D correspond to the same part of the wafer. In other words, reference Figures 2C to 2D , the image generated by photographing the first part of the wafer by an SEM device is the first reference image 200C, and the image generated by photographing this first part by an optical scanner device is the second reference image 200D. In various embodiments, the first reference image 200C and the second reference image 200D are also referred to as a reference image pair. In some embodiments, one or more reference image pairs respectively correspond to one or more parts of the wafer and are used to train the image processing program 111, as described in more detail below with reference to Figures 3 to 5B which will be described in more detail.

[0079] In the operation of defect inspection, first, the wafer is photographed by an optical scanner device to generate a defect map image (not shown). The defect map image includes a plurality of patterns, and each of these patterns is an image corresponding to a part of the wafer. In some embodiments, reference Figures 1 to 2A , the pattern in the defect map image is enlarged to be a defect image 132 as shown in the defect image 200A.

[0080] In addition, in some embodiments, the patterns in the defect map image are images determined by the optical scanner device to have defects. Specifically, by the optical scanner device, in the defect map image, the patterns corresponding to adjacent multiple die and / or cells in the wafer are compared to determine which pattern is different from the others. The pattern different from other adjacent patterns is determined to have a defect. Therefore, the die and / or cell corresponding to this pattern has a defect. In various embodiments, a cell is a repeated part in the wafer and is included in at least one die.

[0081] Next, in the operation of defect inspection, reference Figures 1 to 2B, after the optical scanner device generates at least one defective image 132, since the defective image 132 is an image with unclear picture quality, it is necessary to further inspect the corresponding part of the wafer. In some embodiments, the trained image processing program 111 processes the defective image 132 to implement the operation of further inspecting the corresponding part of the wafer. The trained image processing program 111 is used to process the defective image 132 and generate a reconstructed image 112 as shown in the reconstructed image 200B. Therefore, based on the clear-quality reconstructed image 112, the image processing system 100 can determine whether the corresponding part of the wafer is indeed defective and / or further analyze this defect.

[0082] In some embodiments, referring to Figures 2A to 2B , the picture quality of the image includes, for example, the contour, brightness, contrast, etc. of the defect pattern 210. In some embodiments, the analyzed defect includes multiple pieces of information, where the information includes, for example, the type, size, position, etc. of the object pattern 220 corresponding to the defect.

[0083] Now refer to Figure 3 . According to some embodiments of the present disclosure, Figure 3 is a schematic diagram of the image processing program 111. Figure 3 The image processing program 111 shown is an embodiment regarding Figure 1 . For ease of understanding, the same reference numerals are used hereinafter to designate Figure 1 similar elements in Figure 1 . The following description refers to the image processing system 100 in

[0084] as well to illustrate the configuration and operation of the image processing program 111. Figure 3 As shown in

[0085] In some embodiments, the image generation model 311 includes an encoder and a decoder. In some other embodiments, the image generation model 311 includes a convolution network and a deconvolution network. The convolution network is used to extract the pattern features of the input image (e.g., the defect image 132). The deconvolution network is used to repeatedly generate an image based on the extracted pattern features to reconstruct an output image (e.g., the reconstructed image 112) having such features.

[0086] In some embodiments, the image discrimination model 312 includes a convolution network. The convolution network is used to determine whether the input image (e.g., the reconstructed image 112) contains pattern features corresponding to the wafer defect. In some other embodiments, the convolution network includes at least one convolution layer, at least one pooling layer, and at least one fully connected layer. In various embodiments, a plurality of convolution layers and a plurality of pooling layers are arranged in sequence, and the last pooling layer is coupled to the fully connected layer. The convolution layer is used to extract the pattern features of the input picture. The pooling layer is used to downsample the data output by the convolution layer to reduce the data and retain the pattern features. The fully connected layer is used to flatten the data output by the pooling layer to further output the result.

[0087] Reference Figures 1 to 3 , the trained image processing program 111 is executed by the processor 11b to calculate the input defect image 132 and further output the reconstructed image 112. The operation of executing the trained image processing program 111 by the processor 11b is as described in the following reference Figures 2A to 3 description.

[0088] When the image processing program 111 is executed, the image generation model 311 is used to generate the reconstructed image 112 based on the input defect image 132. In other words, the image generation model 311 converts the defect image 200A into the reconstructed image 200B. In some embodiments, the image generation model 311 is implemented by an image-to-image translation model.

[0089] When generating the reconstructed image 112, the image discrimination model 312 is used to determine whether the reconstructed image 112 contains at least one object pattern. In addition, the image discrimination model 312 is also used to determine whether this at least one object pattern is associated with a part of the wafer photographed under a certain condition.

[0090] In some embodiments, the reconstructed image 112 is equivalent to Figure 2B the reconstructed image 200B of Figure 2B . In this configuration, the object pattern refers to the object pattern 220, which corresponds to the defect of the wafer. In some embodiments, this condition refers to the first condition referred to in Figure 2C Figure 2C . In addition, in some embodiments, this condition is represented by the parameters of the object pattern 220. The parameters include, for example, grayscale values, contour coefficients, etc. This condition means that the object pattern 220 is a pattern captured by an SEM device.

[0091] In some embodiments, in the Figure 3 configuration of Figure 3 , the image discrimination model 312 is used to determine whether the reconstructed image 112 generated by the image generation model 311 is sufficiently similar to the image actually captured by the SEM device. For illustration, as Figures 2A to 3 shown in Figures 2A to 3 , the image discrimination model 312 compares the reconstructed image 200B and the first reference image 200C to determine whether the reconstructed image 200B is similar to the first reference image 200C.

[0092] When the reconstructed image 112 contains at least one object pattern, and this object pattern is associated with a part of the wafer captured under a certain condition, the image discrimination model 312 is used to output the reconstructed image 112. For illustration, as Figures 2A to 3 shown in Figures 2A to 3 , the image discrimination model 312 considers that the reconstructed image 200B contains the object pattern 220, and considers that the various parameters of the object pattern 220 conform to the image parameters captured by the SEM device. Therefore, the image discrimination model 312 outputs the reconstructed image 112 as shown in the reconstructed image 200B.

[0093] In some embodiments, even though the reconstructed image 112 is not an image captured by the SEM device, but is simulated by the image generation model 311 to be an image captured by the SEM device, the reconstructed image 112 is still sufficiently similar to the image captured by the SEM device. For example, referring to Figures 2B to 2C Figures 2B to 2C , the reconstructed image 112 as shown in the reconstructed image 200B is sufficiently similar to the first reference image 200C actually captured by the SEM device.

[0094] For illustration purposes, some embodiments of the present disclosure provide the Figure 3 configuration of the image processing program 111 of Figure 3 , but not limited thereto. Any configurations for implementing the image processing program 111 in Figure 3 Figure 3 are within the scope contemplated by some embodiments of the present disclosure.

[0095] Now refer to Figure 4 . According to some embodiments of the present disclosure, Figure 4 Figure 4 is a flowchart of the image processing method 400. In some embodiments, the image processing method 400 is Figure 1Operation flowchart of the image processing system 100.

[0096] As Figure 4 shown, the image processing method 400 includes operations S410, S420, S430, and S440. The following refers to Figure 1 the image processing system 100 in Figure 4 for an illustration of the method 400, which includes exemplary operations. However, Figure 4 the operations in

[0097] Operation S410, the image processing program is trained by the processor. For illustration, as Figure 1 shown, the executable instructions are executed by the processor 11b to train the image processing program 111. In some embodiments, according to a pair of reference images (such as Figures 2C to 2D the first reference image 200C and the second reference image 200D discussed), the image processing program 111 calculates the pair of reference images to implement the training operation, as described in more detail below with reference to Figures 5A to 5B the operations.

[0098] In operation S420, at least one defect image of the wafer taken by the optical imaging device is obtained. For illustration, as Figure 1 shown, the second electronic device 130 takes a picture of the wafer and generates a defect image 132. The processor 11b accesses the defect image 132. In other words, referring to Figure 2A the defect image 200A of the wafer taken by the optical scanner device is obtained.

[0099] In operation S430, a reconstructed image is generated based on the defect image by the trained image processing program. For illustration, as Figure 1 and Figure 3 shown, the trained image processing program 111 and the defect image 132 are accessed by the processor 11b. In addition, the processor 11b executes executable instructions to cause the image generation model 311 in the trained image processing program 111 to process the defect image 132 to generate a reconstructed image 112. In other words, referring to Figures 2A to 3 when the image processing program 111 is executed, the image generation model 311 converts the defect image 200A into a reconstructed image 200B.

[0100] In some embodiments, operation S430 further includes the following operations. By the trained image processing program, based on at least one reference image, it is determined whether the reconstructed image contains at least one object pattern corresponding to the wafer. For illustration, as Figure 1 and Figure 3As shown, the trained image processing program 111 is accessed by the processor 11b. In addition, executable instructions are executed through the processor 11b so that the image discrimination model 312 in the trained image processing program 111 uses the first reference image 121 (i.e., the first reference image 200C) as a comparison basis to determine whether the reconstructed image 112 (i.e., the reconstructed image 200B) contains an object pattern 220 as shown in Figure 2B or an object pattern 230 as shown in Figure 2C . In other words, referring to Figures 2A to 3 , the trained image discrimination model 312 determines whether the reconstructed image 200B is sufficiently similar to the first reference image 200C based on the first reference image 200C. In various embodiments, as shown in Figure 1 and Figure 3 , the trained image discrimination model 312 compares the first reference image 121 and the reconstructed image 112 to determine whether the reconstructed image 112 is an image generated by an SEM device.

[0101] In operation S440, when the reconstructed image contains at least one object pattern, the reconstructed image is output through the trained image processing program. For illustration, as shown in Figure 1 and Figure 3 , the trained image processing program 111 is accessed by the processor 11b. In addition, executable instructions are executed through the processor 11b so that the image discrimination model 312 in the trained image processing program 111 processes the reconstructed image 112. When the image discrimination model 312 determines that the reconstructed image 112 contains at least one object pattern 220 (shown in Figure 2B ), the reconstructed image 112 is output. In other words, referring to Figures 2A to 3 , the image discrimination model 312 considers that the reconstructed image 200B is an image generated by an SEM device, and thus outputs the reconstructed image 200B.

[0102] In some methods, in the operation of defect inspection, an electron beam imaging device is used to further inspect the defect image to determine or analyze the defects of the wafer. Since an electron beam imaging device (e.g., an SEM device) requires a large amount of manpower and time to generate relatively clear images, the efficiency of the defect inspection operation is low. In addition, the images taken by the SEM device may be unclear images, such as containing electron beam defocus, and the pattern corresponding to the defect cannot be compared with the defect pattern in the defect image.

[0103] Compared with the above methods, in some embodiments of the present disclosure, for example, referring to Figure 1, the trained image processing program 111 is executed by the processor 11b to further automatically inspect the defective image 132. Since the trained image processing program 111 is a processing program, defect inspection can greatly reduce the time and labor of defect inspection operations. In addition, compared with the above methods, defect inspection can generate a reconstructed image 112 that excludes out-of-focus or pattern-shifted images that cannot be compared. On the other hand, in some embodiments of the present disclosure, defect inspection can inspect each defective image 132, omit the operation of taking pictures using the SEM device, improve the wafer per hour (WPH) of on-line wafer inspection, and shorten the downtime of the machine for manufacturing wafers.

[0104] Now refer to Figures 5A to 5B . According to some embodiments of the present disclosure, Figures 5A to 5B is a flowchart of methods 500A and 500B for training an image processing program corresponding to Figure 4 the image processing method 400. The following refers to Figure 1 the image processing system 100 in Figure 3 and Figures 5A to 5B the image processing program 111 in Figures 5A to 5B for illustrations of methods 500A and 500B of Figures 5A to 5B including exemplary operations. However, Figure 1 the operations in

[0105] are not necessarily executed in the order shown in the figures. In other words, within the scope of the concepts of various embodiments of the present disclosure, operations can be appropriately added, replaced, reordered, and / or deleted. In various embodiments, Figure 5A the operations in Figure 3 are further executed by accessing a processor (e.g.,

[0106] the processor 11b shown in

[0107] Figure 1 ), so the discussion of the following operations will not be repeated.

[0105] As Figure 5A shown, the method 500A for training an image processing program includes operations S511, S512, S513, S514, S515, and S516. In some embodiments, the method 500A is applied to training Figure 3 the image discrimination model 312 in

[0106] In some embodiments, at least one algorithm is applied to train the image discrimination model 312. The algorithms include, for example, the Yolo algorithm, the single shot multibox detection (SSD) algorithm, or the regions with convolutional neural network (R-CNN) algorithm, etc.

[0107] In operation S511, the image discrimination model receives a first reference image and a second reference image of the wafer. In some embodiments, the first reference image and the second reference image are images generated by devices using different imaging techniques and are images of the same location on the wafer. For illustration, as Figure 1 shown, the image discrimination model 312 receives the first reference image 121 from the first electronic device 120 and the second reference image 131 from the second electronic device 130. In other words, referring to Figures 2C to 3 , the first reference image 200C captured by the SEM device and the second reference image 200D captured by the optical scanner device are input into the image generation model 311. The first reference image 200C and the second reference image 200D are images of the same part of the wafer.

[0108] In operation S512, the image discrimination model compares the first reference image and the second reference image and generates a comparison value. In some embodiments, the image discrimination model 312 compares whether multiple first reference images are similar to each other and determines whether the above-mentioned first reference images are similar to each other to generate corresponding comparison values. In some other embodiments, the image discrimination model 312 compares whether multiple second reference images are similar to each other and determines whether the above-mentioned second reference images are similar to each other to generate corresponding comparison values.

[0109] In some embodiments, in operation S512, the image discrimination model compares two reference images to determine whether one of these reference images contains at least one object pattern. This object pattern is the reference object pattern 230 as shown in Figure 2C . In some other embodiments, in operation S512, the image discrimination model 312 compares two reference images to further determine whether this object pattern is associated with the wafer captured under specific conditions. In some embodiments, this specific condition is referred to as the first condition as described in Figure 2C . In some embodiments, this specific condition is represented by parameters of the image, where the parameters include, for example, grayscale values, contour coefficients, or similar parameters. This condition means that this image is a pattern captured by the SEM device.

[0110] In some embodiments, the comparison value is one of multiple parameters output according to the loss function in the GAN model. In various embodiments, the comparison value is used to determine whether two images are similar enough. These two images refer to the first and second reference images, multiple first reference images, or multiple second reference images. For illustration, as Figure 1As shown, the image discrimination model 312 compares the first reference image 121 and the second reference image 131, compares multiple first reference images 121 with each other, and / or compares multiple second reference images 131 with each other. In other words, in various embodiments, the reference Figures 2C to 2D , the first reference image 200C and the second reference image 200D are compared with each other, and a comparison result is generated, where the comparison result indicates whether they are similar enough. In other embodiments, multiple images similar to the first reference image 200C or multiple images similar to the second reference image 200D (not shown) are compared with each other, and a comparison result is generated. In some embodiments, the comparison result is represented by the output value of a loss function.

[0111] In operation S513, the image discrimination model determines whether the comparison value is less than a first critical value according to the comparison value. In some embodiments, the first critical value is a parameter set according to the loss function in the GAN model.

[0112] When the comparison value is not less than the first critical value, operation S514 is executed. When the comparison value is less than the first critical value, operation S515 is executed.

[0113] In some embodiments, if there are more difference points between two images, the value output by the loss function (i.e., the comparison value) in operation S512 is higher. For example: reference Figures 2A to 2D , there are multiple difference points between the first reference image 200C and the second reference image 200D. Conversely, if there are fewer difference points between two images, the value output by the loss function in operation S512 is lower. For example: reference Figures 2A to 2D , there are fewer difference points between the first reference image 200C and the second reference image (not shown) similar to the reconstructed image 200B.

[0114] In some embodiments, if the first critical value is set to a lower value, when the comparison value is not less than the first critical value, it indicates that there are still enough difference points (i.e., not similar enough) to distinguish the two images. Conversely, when the comparison value is less than the first critical value, it indicates that the two images are similar enough.

[0115] In operation S514, when the comparison value is greater than or equal to the first critical value, the image discrimination model is updated according to the comparison value. The updated image discrimination model has better discrimination ability to distinguish the similarities and differences between two images. The updated image discrimination model continues to be trained, and operation S512 is continued.

[0116] In operation S515, when the comparison value is less than the first critical value, the training of the image discrimination model is completed. The trained image discrimination model is used to discriminate whether two input images are similar enough. For example: reference Figures 2A to 3, the trained image discrimination model 312 is used to compare any input image with the first reference image 200C to determine whether the input image is similar enough to the first reference image 200C.

[0117] The trained image discrimination model is used to assist in training the image generation model and perform operation S516. In operation S516, operations S517 and S518 are included. Operations S517 and S518 are discussed together with the following reference Figure 5B discussed together.

[0118] As Figure 5B shown, the method 500B for training the image processing program includes operations S521, S522, S523, S524, and S525. In some embodiments, the method 500B is applied to train the Figure 3 image generation model 311 therein.

[0119] In some embodiments, at least one algorithm is applied to train the image generation model 311. The algorithms may include, for example, the U-net algorithm, the GAN algorithm, or the autoencoder algorithm or similar algorithms.

[0120] In operation S521, the image generation model receives the first reference image and the second reference image of the wafer and generates a training reconstruction image based on the first reference image and the second reference image.

[0121] In some embodiments, the first and second reference images correspond to the first and second reference images in operation S511. For illustration, as Figure 1 and Figure 3 shown, the image generation model 311 generates a training reconstruction image (not shown) based on the first reference image 121 and the second reference image 131. In other words, in some embodiments, referring to Figures 2A to 3 , the image generation model 311 converts the second reference image 200D into a training reconstruction image similar to the defect image 200A or the reconstruction image 200B. In some other embodiments, referring to Figures 2A to 3 , the image generation model 311 converts the first reference image 200C into a training reconstruction image similar to the defect image 200A or the reconstruction image 200B.

[0122] Returning to Figure 5A the operations S517 and S518 shown.

[0123] In operation S517, the trained image discrimination model receives the training reconstruction image generated by the image generation model in training in operation S521.

[0124] In operation S518, the trained image discrimination model generates a weight value based on the training reconstruction image to determine whether the image output by the image generation model is similar to the first reference image.

[0125] In some embodiments, since the trained image discrimination model has the function of discriminating whether two images are similar enough, the trained image discrimination model is used to compare the training reconstruction image and the first reference image and generate a comparison result, where this comparison result indicates whether they are similar. In some embodiments, the comparison result is represented by the output value of a loss function and is also referred to as a weight value. In other words, referring to Figures 2A to 3 , the trained image discrimination model 312 compares the training reconstruction image generated by the image generation model 311 with the first reference image 200C to determine whether they are similar and represents it with a weight value.

[0126] Return to Figure 5B the operation S522 shown.

[0127] In operation S522, the image generation model receives the weight value generated by the trained image discrimination model.

[0128] In operation S523, the image generation model determines whether the weight value is less than a second critical value according to the weight value. In some embodiments, the second critical value is a parameter set according to the loss function in the GAN model. In various embodiments, the second critical value is an alternative embodiment of the first critical value.

[0129] When the weight value is not less than the second critical value, operation S524 is executed. When the weight value is less than the second critical value, operation S525 is executed.

[0130] In some embodiments, if there are more difference points between the training reconstruction image and the first reference image, the value output in operation S518 (i.e., the weight value) is higher. For example: referring to Figures 2A to 2D , there are multiple difference points between the training reconstruction image similar to the second reference image 200D and the first reference image 200C. In other words, referring to Figures 2A to 3 , the trained image discrimination model 312 believes that the training reconstruction image and the first reference image 200C are not similar enough. In another way of explanation, the trained image discrimination model 312 believes that the training reconstruction image is not an image generated by the SEM device.

[0131] In some embodiments, if there are fewer difference points between the training reconstruction image and the first reference image, the value output in operation S518 (i.e., the weight value) is lower. For example: referring to Figures 2A to 2D , the training reconstruction image corresponds to the reconstruction image 200B and is similar to the first reference image 200C. In other words, referring to Figures 2A to 3, the trained image discrimination model 312 determines that the training reconstructed image is sufficiently similar to the first reference image 200C. In other words, the trained image discrimination model 312 is misled by the image generation model 311 and determines that the training reconstructed image is an image captured by the SEM device.

[0132] In operation S524, according to the weight value, when the weight value is greater than or equal to the second threshold value, the image generation model is updated. The updated image generation model is used to convert the input image into an image more similar to the first reference image to simulate an image captured by the SEM device. The updated image generation model then continues to execute operation S522 to continue the training.

[0133] In operation S525, when the weight value is less than the second threshold value, the training of the image generation model is completed. The trained image generation model is used to convert the input image into an image similar to an image captured by the SEM device, where this image is used to make the trained image discrimination model believe it is an image captured by the SEM device. For example, refer to Figures 2A to 3 , the trained image generation model 311 is used to convert the input defective image 200A into a reconstructed image 200B. In addition, the trained image discrimination model 312 determines that the reconstructed image 200B is an image captured by the SEM device.

[0134] In some embodiments, methods 500A and 500B are alternately executed N times, where N is a positive integer. For example, when operation S525 is completed, the originally trained image discrimination model then re-executes method 500A to retrain the image discrimination model again. During the second training of the image discrimination model, the first and second reference images input in operation S511 are changed to the training reconstructed images generated by the trained image generation model. Similarly, when the image discrimination model completes the second training, it then re-executes method 500B to retrain the image generation model, and so on.

[0135] In some embodiments, when method 500A is executed, the image generation model is considered to be trained, and only the image discrimination model is updated. In some other embodiments, when method 500B is executed, the image discrimination model is considered to be trained, and only the image generation model is updated. In various embodiments, if methods 500A and 500B are executed multiple times (i.e., N is greater than 1), at least one of the first threshold value or the second threshold value is different from each other in each corresponding method 500A and 500B.

[0136] Now refer to Figure 6 and Figure 7 . According to some embodiments of the present disclosure, Figure 6is a flowchart of an image processing method 600. In some embodiments, the image processing method 600 is Figure 1 the operating flowchart of the image processing system 100. In various embodiments, the image processing method 600 is Figure 4 an alternative embodiment of the image processing method 400. According to some embodiments of the present disclosure, Figure 7 is a schematic diagram corresponding to Figure 6 the operation of classifying defective images of the image processing method.

[0137] As Figure 6 shown, the image processing method 600 includes operations S610, S620, S630, S640, and S650. The following refers to Figure 1 the image processing system 100 in Figure 6 for an illustration of the method 600 including exemplary operations. However, Figure 6 the operations in

[0138] operation S610, the image processing program is trained by a processor. In some embodiments, operation S610 corresponds to Figure 4 operation S410 in

[0139] operation S620, a plurality of defective images of the wafer taken by an optical imaging device are obtained. In some embodiments, operation S620 corresponds to Figure 4 operation S420 in

[0140] operation S630, the processor classifies the defective images into a plurality of image groups according to the attributes of the defective images. For illustration, as Figure 1 shown, a plurality of defective images 132 are input into the memory 11a of the image processing device 110. The processor 11b executes executable instructions to divide the defective images 132 into a plurality of groups according to the attributes of the defective images 132 corresponding to the wafer.

[0141] In some embodiments, for the image processing device 110, the defective images 132 are also referred to as input images. In various embodiments, the attribute refers to the circuit attribute of the wafer part corresponding to the defective image 132. For example: which type of circuit in the integrated circuit this part belongs to. The circuit type includes, for example, a memory circuit or other logic circuits, etc. In other words, referring to Figure 7, multiple defective images are classified into a first image group and a second image group. The first image group is represented by block 711, and the defective images 132 included in the first image group belong to the memory circuit. The second image group is represented by block 712, and the defective images 132 included in the second image group belong to the logic circuit.

[0142] In some embodiments, operation S630 further includes the following operations. According to multiple features of the attributes of the defective images, the defective images are classified one by one in the corresponding image group. For illustration, as Figure 1 and Figure 7 shown, by executing executable instructions through the processor 11b, in the first image group (block 711), according to multiple features of the defective images 132, these defective images 132 are classified into multiple sub-image groups. These sub-image groups are represented by other blocks such as including blocks 721 and 722. In addition, in the sub-image groups (blocks 721 and 722), according to multiple features of these defective images 132 respectively, these defective images 132 are continuously classified into image groups of other branches. These image groups of branches are represented by other blocks such as including block 731.

[0143] In some embodiments, the features of the defective images include the optical parameters of the defective images. The optical parameters include, for example, polarity, brightness, contrast, etc. In some embodiments, the features of the defective images include the graphic parameters of the defect pattern and the background pattern in the defective images. The graphic parameters include, for example, area ratio, eccentricity, and their relative positions, etc.

[0144] In some embodiments, the first and second image groups (blocks 711 and 712) are respectively classified into multiple image groups of branches with different features. For example, referring to Figure 7 , one of the first image groups (block 711) (block 731) has attributes and features such as a memory circuit, a polarity higher than the first reference value, and a defect pattern shape approximately circular, etc. One of the second image groups (block 712) (blocks 732 or 733) has attributes and features such as a power supply circuit and a brightness between the second and third reference values, etc.

[0145] In some embodiments, the number of times of classifying the attributes and features of the defective images in operation S630 is greater than 4 times, thereby obtaining multiple image groups with similar properties to perform subsequent operations. In some embodiments, the number of times of classifying the attributes and features of the defective images in operation S630 is less than 7 times, thereby efficiently obtaining multiple image groups with similar properties to perform subsequent operations.

[0146] In some embodiments, operation S610 is performed after operation S630.

[0147] In some embodiments, operation S630 further includes the following operations. After classifying the defective images into multiple image groups, the processor executes executable instructions to train an image processing program based on the defective images in each image group and the images captured by the corresponding SEM device. In other words, for each of the image groups, the defective images with the same attributes and characteristics and the images captured by the corresponding SEM device are regarded as a reference image pair. The image processing program calculates the reference image pairs to achieve the training operation. Therefore, the trained image processing program is used to perform image processing operations on the defective images in this image group. In various embodiments, in this configuration, the image processing system includes multiple image processing programs, and these image processing programs are respectively used to process the defective images in different image groups. For illustration, as Figure 1 shown, the image processing system 100 includes multiple image processing programs 111, and the image processing programs 111 respectively correspond to defective images 132 with different attributes and characteristics.

[0148] In operation S640, the processor selects, in one of the image groups, an image processing program corresponding to these features based on multiple features of the attributes of the defective images. For illustration, as Figure 1 and Figure 7 shown, the processor 11b accesses one of the image groups (e.g., block 711). The processor 11b executes executable instructions to select, based on multiple features of the defective image 132 in one of the image groups (e.g., block 711), to access the defective images 132 with these features (e.g., block 731) and the image processing program 111 corresponding to these features.

[0149] In operation S650, the processor uses the selected image processing program to process the defective image to output a corresponding reconstructed image. For illustration, as Figure 1 and Figure 7 shown, if the processor 11b selects the image processing program 111 corresponding to block 731 in operation S640, the processor 11b executes executable instructions to access this image processing program 111 and cause it to process the defective image 132 included in block 731, and output the corresponding reconstructed image 112.

[0150] In some embodiments, operation S650 corresponds to Figure 4 the operations S430 to S440 in, and the same parts are not described herein again.

[0151] Now refer to Figures 8A to 8C . According to some embodiments of the present disclosure, Figures 8A to 8C corresponds to Figure 7Schematic diagrams of defective images 800A, 800B, and 800C of the classification results. In some embodiments, defective images 800A, 800B, and 800C are Figure 1 Examples of the defective image 132 shown. In various embodiments, defective images 800A, 800B, and 800C are Figure 2A Alternative embodiments of the defective image 200A shown.

[0152] As Figure 8A shown, in some embodiments, defective image 800A is multiple images corresponding to different parts of the wafer. In various embodiments, defective image 800A is a defective image in an image group, for example: Figure 7 The image group represented by square 731 in.

[0153] In some embodiments, for defective image 800A, the attributes include that the circuit type is a logic circuit. The features include that the defect pattern is bright with a dark flash in the middle and has a lower polarity.

[0154] As Figure 8B shown, in some embodiments, defective image 800B is an alternative embodiment of defective image 800A. In various embodiments, defective image 800B is a defective image in an image group, for example: Figure 7 The image group represented by square 732 in.

[0155] In some embodiments, for defective image 800B, the attributes include the circuit type, and this circuit type can belong to a memory circuit. The features include that the defect pattern is black with a bright flash and is located behind the background pattern.

[0156] As Figure 8C shown, in some embodiments, defective image 800C is an alternative embodiment of defective image 800A or 800B. In various embodiments, defective image 800C is a defective image in an image group, for example: Figure 7 The image group represented by square 733 in.

[0157] In some embodiments, for defective image 800C, the attributes include that the circuit type is a logic circuit. The features include that the defect pattern is a whole bright flash and has a higher polarity.

[0158] In some embodiments, an image processing method is disclosed. The image processing method includes the following steps: obtaining a defective image of a wafer; processing this defective image and generating a reconstructed image; and when this reconstructed image includes at least one object pattern, outputting this reconstructed image. This at least one object pattern corresponds to a part of this wafer.

[0159] In some embodiments, the image processing method further comprises the following steps: training an image processing program using a plurality of reference images. The trained image processing program is used to process the defective image that does not have the at least one object pattern to generate and output the reconstructed image.

[0160] In some embodiments, these reference images correspond to multiple parts of the wafer. These reference images include a plurality of first reference images and a plurality of second reference images. These first reference images and the reconstructed image are associated with the wafer taken under a first condition. These second reference images and the defective image are associated with the wafer taken under a second condition different from the first condition.

[0161] In some embodiments, the image processing method further comprises the following steps: training an image generation model using a plurality of reference images. These reference images correspond to multiple parts of the wafer. The trained image generation model is used to convert the defective image into the reconstructed image.

[0162] In some embodiments, the image processing method further comprises the following steps: capturing the wafer by a plurality of electronic devices to obtain a plurality of reference images of the wafer; and training an image discrimination model using these reference images. The trained image discrimination model is used to compare the reconstructed image and these reference images to determine whether the reconstructed image contains the at least one object pattern, and to determine whether the at least one object pattern is associated with the wafer taken under a condition.

[0163] In some embodiments, the image processing method further comprises the following steps: obtaining a plurality of input images including the defective image; classifying these input images into a plurality of image groups according to a plurality of attributes of these input images, wherein these input images correspond to multiple parts of the wafer.

[0164] In some embodiments, the image processing method further comprises the following steps: selectively using the trained image processing program to process the defective image according to at least one attribute of the defective image to output the reconstructed image. The image processing program is associated with at least one attribute of the defective image.

[0165] In some embodiments, a non-transitory computer-readable medium is also disclosed. The non-transitory computer-readable medium includes executable instructions for implementing an image processing method by a processor. The image processing method includes the following steps: capturing a defect image, where the defect image includes a defect pattern corresponding to a wafer; and generating a reconstructed image based on the defect image and a plurality of first reference images, where the reconstructed image includes at least one object pattern corresponding to the wafer, and the first reference images include a plurality of patterns corresponding to the wafer. At least one object pattern and the defect pattern correspond to the same part of the wafer, and the at least one object pattern is different from the defect pattern.

[0166] In some embodiments, the image processing method further includes the following steps: training an image generation model using the first reference images and a plurality of second reference images corresponding to the first reference images. The trained image generation model is used to convert the defect image into the reconstructed image.

[0167] In some embodiments, the image processing method further includes the following steps: photographing the wafer by a first electronic device to obtain the first reference images; and photographing the wafer by a second electronic device to obtain the second reference images and the defect image.

[0168] In some embodiments, the operation of training the image generation model includes: generating a plurality of training reconstructed images based on the first reference images and the second reference images; receiving a plurality of weight values, where the weight values represent comparison results of the training reconstructed images and the first reference images; and updating the image generation model based on the weight values.

[0169] In some embodiments, the image processing method further includes the following steps: training an image discrimination model using the first reference images and a plurality of second reference images corresponding to the first reference images. The trained image discrimination model is used to generate at least one weight value to the image generation model to determine whether at least one image output by the image generation model is similar to the first reference images.

[0170] In some embodiments, the operation of training the image discrimination model includes: comparing the first reference images and the second reference images to discriminate the first reference images as being related to the wafer photographed under a condition.

[0171] In some embodiments, the image processing method further includes the following steps: obtaining a plurality of input images of the wafer photographed by an electronic device, where the input images include the defect image; and classifying the input images into a plurality of image groups according to a plurality of attributes of the input images.

[0172] In some embodiments, in each of these groups of images, based on a plurality of features in these input images, a plurality of image generation models and a plurality of image discrimination models corresponding to these features are selected to process these input images respectively to generate a plurality of output images. These output images include this reconstructed image.

[0173] In some embodiments, these first reference images include at least one reference object pattern corresponding to a wafer, and this at least one reference object pattern is associated with the above-mentioned at least one object pattern.

[0174] In some embodiments, an image processing system is also disclosed. The image processing system includes a memory and a processor. The memory is used to store an image processing program. The processor is coupled to this memory. This processor is used to access this image processing program in this memory to perform the following steps: training this image processing program using a plurality of reference images; and processing a defective image through this trained image processing program to generate and output a reconstructed image. This reconstructed image includes at least one object pattern corresponding to a part of the wafer, and these reference images include a plurality of reference object patterns corresponding to a plurality of parts of this wafer.

[0175] In some embodiments, the image processing program includes an image generation model and an image discrimination model. This processor is further used to access this image processing program in this memory to perform the following steps: converting this defective image into this reconstructed image using this image generation model; and using this image discrimination model to receive this reconstructed image and comparing this reconstructed image with these reference images to determine whether this at least one object pattern is associated with the wafer photographed under a condition.

[0176] In some embodiments, the processor is further used to access this image processing program in this memory to select a trained image processing program corresponding to at least one feature of the defective image according to at least one feature of the defective image and process the defective image.

[0177] In some embodiments, the image processing system further includes a first electronic device and a second electronic device. The first electronic device is coupled to the memory and the processor and is used to photograph the wafer to obtain a first part of these reference images. The second electronic device is coupled to the memory and the processor and is used to photograph the wafer to obtain a second part of these reference images and a defective image.

[0178] The foregoing outlines the features of several embodiments, enabling those skilled in the art to better understand aspects of some embodiments of the present disclosure. Those skilled in the art should understand that these artisans can readily use some embodiments of the present disclosure as a basis for designing or modifying other processes and structures for achieving the same purposes and / or attaining the same advantages as the embodiments described herein. Those skilled in the art should also recognize that these equivalent constructs do not depart from the spirit and scope of some embodiments of the present disclosure, and that these artisans can make various changes, substitutions, and alterations herein without departing from the spirit and scope of some embodiments of the present disclosure.

Claims

1. An image processing method, characterized in that, Comprising: Obtaining a defect image of a wafer, the defect image being a photograph generated by an electronic device photographing the wafer; Processing the defect image through an image generation model to generate a reconstructed image, the reconstructed image being an image of the wafer simulated by electron beam extraction; and When the reconstructed image includes at least one object pattern, outputting the reconstructed image, wherein the at least one object pattern corresponds to a part of the wafer, the defect image includes a defect pattern corresponding to the part of the wafer, and the at least one object pattern is a pattern of the part simulated by the image generation model.

2. The image processing method according to claim 1, wherein, Further comprising: Training an image processing program using a plurality of reference images, wherein the trained image processing program is used to process the defect image without the at least one object pattern to generate and output the reconstructed image.

3. The image processing method according to claim 2, wherein wherein the reference images correspond to multiple parts of the wafer, and the reference images include: Multiple first reference images, wherein the first reference images and the reconstructed image are associated with the wafer photographed under a first condition; and Multiple second reference images, wherein the second reference images and the defect image are associated with the wafer photographed under a second condition different from the first condition.

4. The image processing method according to claim 1, wherein Further comprising: Training the image generation model using a plurality of reference images, wherein the reference images correspond to multiple parts of the wafer, and the trained image generation model is used to convert the defect image into the reconstructed image.

5. The image processing method according to claim 1, wherein Further comprising: Photographing the wafer through a plurality of electronic devices to obtain a plurality of reference images of the wafer; and Training an image discrimination model using the reference images, wherein the trained image discrimination model is used to compare the reconstructed image and the reference images to determine whether the reconstructed image includes the at least one object pattern, and to determine whether the at least one object pattern is associated with the wafer photographed under a condition.

6. The image processing method according to claim 1, wherein Further comprising: Obtaining a plurality of input images including the defect image; and Classifying the input images into a plurality of image groups according to multiple attributes of the input images; wherein the input images correspond to multiple parts of the wafer.

7. The image processing method according to claim 1, wherein Further comprising: Selectively processing the defect image using a trained image processing program according to at least one attribute of the defect image to output the reconstructed image; wherein the image processing program is associated with the at least one attribute of the defect image.

8. A non-transitory computer-readable medium includes executable instructions for implementing an image processing method by a processor, characterized in that, The image processing method comprises: Extracting a defect image, wherein the defect image includes a defect pattern corresponding to a wafer; and Generating a reconstructed image through an image generation model according to the defect image and a plurality of first reference images, wherein the reconstructed image includes at least one object pattern corresponding to the wafer, and the first reference images include a plurality of patterns corresponding to the wafer, wherein the at least one object pattern and the defect pattern correspond to the same part of the wafer, and the at least one object pattern is different from the defect pattern, the defect image is a photograph generated by an electronic device photographing the wafer, and the reconstructed image is an image of the wafer simulated by electron beam extraction.

9. The non-transitory computer-readable medium according to claim 8, wherein, wherein the image processing method further comprises: Train the image generation model by using the first reference images and a plurality of second reference images corresponding to the first reference images. The trained image generation model is used to convert the defective image into the reconstructed image.

10. The non-transitory computer-readable medium according to claim 9, wherein The image processing method further includes: Taking pictures of the wafer by a first electronic device to obtain the first reference images; and Taking pictures of the wafer by a second electronic device to obtain the second reference images and the defective image.

11. The non-transitory computer-readable medium according to claim 9, wherein The operation of training the image generation model includes: Generating a plurality of training reconstructed images according to the first reference images and the second reference images; Receiving a plurality of weight values, where the weight values represent a plurality of comparison results between the training reconstructed images and the first reference images; and Updating the image generation model according to the weight values.

12. The non-transitory computer-readable medium according to claim 8, wherein The image processing method further includes: Training an image discrimination model by using the first reference images and a plurality of second reference images corresponding to the first reference images. The trained image discrimination model is used to generate at least one weight value to the image generation model to determine whether at least one image output by the image generation model is similar to the first reference images.

13. The non-transitory computer-readable medium according to claim 12, wherein The operation of training the image discrimination model includes: Comparing the first reference images and the second reference images to identify whether the first reference images are related to the wafer taken under a condition.

14. The non-transitory computer-readable medium according to claim 8, wherein The image processing method further includes: Obtaining a plurality of input images of the wafer taken by the electronic device, where the input images include the defective image; and Classifying the input images into a plurality of image groups according to a plurality of attributes of the input images.

15. The non-transitory computer-readable medium according to claim 14, wherein The image processing method further includes: In each of the image groups, according to a plurality of features in the input images, selecting a plurality of image generation models and a plurality of image discrimination models corresponding to the features, and respectively processing the input images to generate a plurality of output images. The output images include the reconstructed image.

16. The non-transitory computer-readable medium according to claim 8, wherein The first reference images include at least one reference object pattern corresponding to the wafer, and the at least one reference object pattern is related to the at least one object pattern.

17. An image processing system, characterized in that, Includes: A memory for storing an image processing program, where the image processing program includes an image generation model; and A processor coupled to the memory, and the processor is used to access the image processing program in the memory to execute the following steps: Comparing whether a plurality of reference images are similar to each other to generate corresponding comparison values to train the image processing program; and Processing a defective image through the trained image processing program to generate and output a reconstructed image. The reconstructed image includes at least one object pattern corresponding to a part of the wafer, and the reference images include a plurality of reference object patterns corresponding to a plurality of parts of the wafer. The defective image is a photo taken by an electronic device of the wafer. The reconstructed image is an image of the wafer simulated by electron beam extraction. The defective image includes a defect pattern corresponding to the part of the wafer. The at least one object pattern is a pattern of the part simulated by the image generation model, and the reference object patterns are the patterns corresponding to the part in the reference images.

18. The image processing system according to claim 17, wherein The image processing program further includes: an image discrimination model, wherein the processor is further configured to access the image processing program in the memory to perform the following steps: using the image generation model, converting the defective image into the reconstructed image; and using the image discrimination model, receiving the reconstructed image and comparing the reconstructed image with the reference images to determine whether the at least one object pattern is associated with the wafer photographed under a condition.

19. The image processing system according to claim 17, wherein wherein the processor is further configured to access the image processing program in the memory to select the trained image processing program corresponding to at least one feature of the defective image according to at least one feature of the defective image to process the defective image.

20. The image processing system according to claim 19, wherein, It further includes: a first electronic device, coupled to the memory and the processor, and configured to photograph the wafer to obtain a first part of the reference images; and a second electronic device, coupled to the memory and the processor, and configured to photograph the wafer to obtain a second part of the reference images and the defective image.

Images