Image recognition method and device, equipment, storage medium and program product

Abstract

The invention discloses an image recognition method and device, equipment, a storage medium and a program product, and relates to the technical field of semiconductors, and the method comprises the steps: determining a pre-created initial classification tree corresponding to the image category of a to-be-recognized image; performing structure adjustment on the initial classification tree to obtain a target classification tree; each leaf node in the target classification tree corresponds to each target feature category needing to be identified in the current to-be-identified image; and identifying a target image feature of the to-be-identified image, and determining a corresponding image identification result based on the target image feature and a target feature category corresponding to each leaf node in the target classification tree. Classification is carried out through the multi-level tree structure, and the number of layers and the structure of the multi-level tree structure can be flexibly changed, so that a more flexible recognition effect is realized for different number and different types of feature categories and different image input requirements which need to be recognized.

Description

Technical Field

[0001] The present application relates to the field of semiconductor technology, and particularly to an image recognition method, device, equipment, storage medium and program product. Background Art

[0002] In the lithography process of semiconductor technology, the pre-designed integrated circuit pattern on the mask will be transferred to multiple semiconductor wafers. Therefore, any defect on the photomask may affect the subsequent wafers during the subsequent lithography process, resulting in the yield problem of the wafers. Therefore, during the manufacturing of the mask and the subsequent use process after mask manufacturing, it is necessary to continuously detect the mask to determine whether there are defects on the mask. The defects at different mask positions have different impacts on the subsequent lithography process. Therefore, in order to better classify serious defects and interference signals and distinguish defect types in the future, a geometric position classification step is required during the lithography process.

[0003] The geometric classification step can classify the geometric features of the neighborhood where each pixel is located in the entire image to be detected during the detection process, and output the geometric position information of the surrounding neighborhood where each pixel is located, which is used as a reference for the geometric position of the pattern where the defect is located during subsequent defect detection. However, due to the existence of multiple detection algorithm modes in the subsequent mask detection process, different detection algorithm modes correspond to different data inputs. For the current geometric classification method, separate additional schemes need to be designed for various input data. While the calculation complexity is high, the low universality of the algorithm and the high subsequent maintenance cost brought by the multi-channel algorithm also become a major problem. In addition, due to the increasing complexity of the current integrated circuit pattern and the increasing number of geometric position categories on the mask pattern, the flexibility of the current geometric position classification mode for geometric position categories still needs to be improved. Summary of the Invention

[0004] The present application discloses an image recognition method, device, equipment, storage medium and program product, which is used to solve the problem of insufficient flexibility in geometric position classification on the mask pattern.

[0005] In a first aspect, the present application provides an image recognition method, including: determining an initial classification tree pre-created corresponding to the image category of the image to be recognized; performing structural adjustment on the initial classification tree to obtain a target classification tree; each leaf node in the target classification tree corresponding to each target feature category to be recognized in the current image to be recognized; recognizing the target image features of the image to be recognized, and determining corresponding image recognition results based on the target image features and the target feature categories corresponding to the leaf nodes in the target classification tree. In this solution, the target features in the image are classified through a multi-level tree structure. Since the number of layers and the structure of the multi-level tree can be flexibly changed, it is possible to more flexibly recognize input images with different numbers and types of categories.

[0006] In a possible implementation manner, the image recognition method further includes: performing image alignment on the initial image based on a preset reference image to obtain an aligned image; using the aligned image as the image to be recognized. In this implementation manner, by performing image alignment on the image to be recognized based on the preset reference image before image recognition, it is possible to ensure that the data offset caused by errors in the data acquisition process between the image to be processed and the reference image is well corrected, which helps to improve the accuracy of image recognition.

[0007] In a possible implementation manner, the initial image is an image to be recognized in a mask detection process, including any one or several of a first image designed by electronic design automation software and including corresponding vertices of a polygon, a second image obtained by rasterizing the corresponding vertices of the polygon, and a third image obtained by imaging the mask. In this implementation manner, the input of image data in the present application can be various types of images. When the input image is of different types, there is no need to adjust the recognition scheme every time, and it is possible to flexibly recognize different types of input images.

[0008] In a possible implementation manner, the performing image alignment on the initial image based on a preset reference image includes: obtaining the initial image and the corresponding preset reference image; performing preprocessing on the initial image and the preset reference image to obtain a preprocessed initial image and a preprocessed reference image; performing alignment on the preprocessed initial image based on the preprocessed reference image; where the preprocessing includes image denoising and / or image gray level matching. In this implementation manner, performing image alignment based on the preprocessed initial image and reference image can better correct the data offset when obtaining the initial image.

[0009] In a possible implementation, before determining the initially created initial classification tree corresponding to the image category of the image to be recognized, it further includes: determining each image category; counting each feature category possessed by the images of each of the image categories to obtain respective sets of feature categories corresponding to each of the image categories; creating respective initial classification trees corresponding to each of the image categories based on the image categories and the sets of feature categories; wherein each level of the initial classification tree is completely filled, and each leaf node of the initial classification tree corresponds to each feature category in the corresponding set of feature categories. In this implementation, by pre-constructing the corresponding maximum classification tree as the initial classification tree based on the feature categories corresponding to each image category, during subsequent image recognition, the structure of the classification tree can be adjusted according to different target features to be recognized to obtain the target classification tree, so as to improve the speed and flexibility of image feature classification.

[0010] In a possible implementation, recognizing the target image feature of the image to be recognized includes: performing binarization processing on the image to be recognized to obtain a binarized image; processing the binarized image to obtain the target image feature of the image to be recognized. In this implementation, before performing edge recognition of the image, binarization processing of the image is first performed to facilitate more convenient and efficient extraction of image features subsequently.

[0011] In a possible implementation, processing the binarized image to obtain the target image feature of the image to be recognized includes: performing edge recognition on the binarized image to obtain a target edge image; recognizing defective edges in the binarized image and the target edge image; determining the first target image feature of the image to be recognized based on the position information corresponding to the defective edges. In this implementation, edge recognition is performed on the binarized image to obtain the edge defects therein. Since the edge defects have higher precision and higher recognition sensitivity, the recognition accuracy of the data after optical proximity correction and sub-resolution assist feature processing can be improved; and since the detection of edge defects has no specific requirements for the input of data, when the types of input image data are different, the efficiency can be further improved.

[0012] In a possible implementation manner, the edge recognition of the binary image to obtain a target edge image includes: performing edge recognition on the binary image to obtain an initial edge image; performing a connectivity test on the edge patterns in the initial edge image, and connecting the target edge lines in the edge patterns based on the test result to obtain a target edge image; wherein, the target edge lines are the discontinuous edge lines in the edge patterns. In this implementation manner, when performing pattern edge recognition on data, due to the influence of image resolution, problems such as discontinuous edges may occur. A connectivity test is performed on the discontinuous edge situation and connected to improve the effect of image edge recognition.

[0013] In a possible implementation manner, the recognition of the defective edges in the binary image and the target edge image includes: performing edge smoothing on the target edge image to obtain the target edge image after edge smoothing; performing defective edge recognition on the binary image and the target edge image after edge smoothing. In this implementation manner, before using edge features for geometric recognition, the image edge is smoothed, avoiding the possible misjudgment caused by the relatively complex pattern edges of the mask due to the addition of techniques such as sub-resolution assist features.

[0014] In a possible implementation manner, the recognition of the defective edges in the binary image and the target edge image includes: detecting each pattern edge with a protrusion or depression in the binary image and the target edge image; determining the defective edges from each of the pattern edges based on the geometric parameters of the pattern edges; wherein, the defective edges are the pattern edges whose geometric parameters meet the preset defective edge conditions. In this implementation manner, through the detection and quantification of edge defects, the recognition accuracy of the data after optical proximity correction and sub-resolution assist feature processing is greatly improved, thereby improving the robustness of the geometric classification result.

[0015] In a possible implementation manner, the determination of the defective edges from each of the pattern edges based on the geometric parameters of the pattern edges includes: performing dimensionality reduction processing on the pattern edges based on the geometric parameters to obtain one-dimensional features for describing the pattern edges; using the one-dimensional features to determine the defective edges from each of the pattern edges. In this implementation manner, through the dimensionality reduction processing of the pattern edges, a refined quantitative description of the changes in the one-dimensional edges can be performed, and the recognition accuracy of the data after optical proximity correction and sub-resolution assist feature processing is greatly improved.

[0016] In a possible implementation manner, the processing of the binary image to obtain the target image features of the image to be recognized includes: performing line scanning on the binary image based on a target number of preset line scanning directions; for each region of the binary image scanned in each preset line scanning direction, generating corresponding minimum rectangle results after segmentation in each direction; and determining the second target image features of the image to be recognized according to the minimum rectangle results after segmentation in each direction. In this implementation manner, by further determining the features within and between the minimum rectangles, the feature group can be expanded, and the robustness of the geometric classification result can be further improved.

[0017] In a possible implementation manner, the determining of the corresponding image recognition result based on the target image features and the target feature categories corresponding to the leaf nodes in the target classification tree includes: matching each of the target image features with each of the target feature categories to obtain corresponding matching results; and generating a geometric feature map of the image to be recognized according to the matching results. In this implementation manner, by performing geometric position classification on the target classification tree obtained by excising nodes from the complete tree, the recognition efficiency can be improved by reducing the number of nodes, that is, reducing operations.

[0018] In a possible implementation manner, the generating of the geometric feature map of the image to be recognized according to the matching results includes: determining the priorities among the target feature categories corresponding to the leaf nodes; and generating the geometric feature map of the image to be recognized based on the priorities and the matching results. In this implementation manner, by generating several feature maps based on the priorities among the target feature categories corresponding to different leaf nodes, the practicability of the geometric feature map can be effectively improved.

[0019] Second aspect, the present application provides an image recognition device, including:

[0020] A classification tree determination module, configured to determine an initial classification tree created in advance corresponding to the image category of the image to be recognized;

[0021] A classification tree adjustment module, configured to perform structural adjustment on the initial classification tree to obtain a target classification tree; each leaf node in the target classification tree corresponds to each target feature category to be recognized in the current image to be recognized;

[0022] A feature recognition module, configured to recognize the target image features of the image to be recognized, and determine the corresponding image recognition result based on the target image features and the target feature categories corresponding to the leaf nodes in the target classification tree.

[0023] In a third aspect, the present application provides an electronic device, which includes a processor and a memory; wherein, the memory is used to store a computer program, and the computer program is loaded and executed by the processor to implement the foregoing image recognition method.

[0024] In a fourth aspect, the present application provides a computer-readable storage medium for storing a computer program, and when the computer program is executed by a processor, the foregoing image recognition method is implemented.

[0025] In a fifth aspect, the present application provides a computer program product, including a computer program / instructions, and when the computer program / instructions are executed by a processor, the foregoing image recognition method is implemented.

[0026] In the present application, first, an initial classification tree corresponding to the image category of the image to be recognized created in advance is determined, and the structure of the initial classification tree is adjusted to obtain a target classification tree, where each leaf node in the target classification tree corresponds to each target feature category to be recognized in the current image to be recognized. Then, the target image features of the image to be recognized are identified, and based on the target image features and the target feature categories corresponding to the leaf nodes in the target classification tree, the corresponding image recognition result is determined. Through the above technical solution, the present application can perform classification through a multi-level tree structure. In this way, since the number of levels and the structure of the multi-level tree can be flexibly changed, a more flexible recognition effect is achieved for different numbers, different types of feature categories to be recognized, and different image input requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0028] Figure 1 It is a result diagram of an image recognition system provided by an embodiment of the present application;

[0029] Figure 2 It is a flowchart of an image recognition method provided by an embodiment of the present application;

[0030] Figure 3 It is a flowchart of an image recognition provided by an embodiment of the present application;

[0031] Figure 4 It is a flowchart of an edge defect geometric position recognition provided by an embodiment of the present application;

[0032] Figure 5A schematic diagram of a multi-level classification tree provided by an embodiment of the present application;

[0033] Figure 6 A flowchart of a specific feature recognition method provided by an embodiment of the present application;

[0034] Figure 7 A schematic diagram of an edge smoothing result provided by an embodiment of the present application;

[0035] Figure 8 A schematic diagram of an edge defect provided by an embodiment of the present application;

[0036] Figure 9 A schematic diagram of a line scan result provided by an embodiment of the present application;

[0037] Figure 10 A schematic diagram of a geometric feature provided by an embodiment of the present application;

[0038] Figure 11 A flowchart of an image recognition provided by an embodiment of the present application;

[0039] Figure 12 A schematic diagram of the structure of an image recognition device provided by an embodiment of the present application

[0040] Figure 13 A structural diagram of an electronic device provided by an embodiment of the present application. Detailed implementation manners

[0041] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present application.

[0042] In the lithography process of semiconductor technology, a geometric position classification step is required to output the geometric position information of the surrounding neighborhood where each pixel is located. However, due to the existence of multiple detection algorithm modes in the subsequent mask detection process, different detection algorithm modes correspond to different data inputs, and the current geometric classification methods require separately designed additional schemes for various input data, with a relatively high computational complexity. The present application can perform classification through a multi-level tree structure. In this way, since the number of layers and the structure of the multi-level tree can be flexibly changed, a more flexible recognition effect is achieved for different numbers, different types of feature categories to be recognized, and different image input requirements.

[0043] It should be noted that this application is mainly applied to the process of detecting a mask by a lithography process based on semiconductor technology to determine whether there are defects on the mask. The applicable scenario can be within the scope of mask optical metrology, and corresponding application processes can also be implemented in other fields that require pixel-level geometric classification. Specifically, as Figure 1 shown, this application mainly includes a detection control system and an image / information acquisition system. After the user issues an instruction to the detection control system first, the detection control system can send an instruction for image acquisition to the image / information acquisition system to obtain an initial image to be recognized returned by the image / information acquisition system, and perform a preset detection process through the detection control system to return the obtained result to the user terminal.

[0044] Embodiment 1

[0045] See Figure 2 shown. An image recognition method disclosed in an embodiment of this application includes:

[0046] Step S11, determining an initial classification tree created in advance corresponding to the image category of the image to be recognized.

[0047] In this embodiment, as Figure 3 shown, when executing the image recognition process by a mask detection device, after data is input from the outside, a data preprocessing process will be performed first. This preprocessing process will perform an overall noise reduction, gray-scale matching, etc. process on the data to be processed and the reference. It can be understood that the above preprocessing process is not limited to noise reduction and gray-scale matching, and can be adjusted accordingly according to the actual image preprocessing requirements. After preprocessing the input image, image alignment will be performed on the data to ensure that the data offset caused by errors in the data acquisition process between the image to be processed for analyzing and judging the presence and location of defects and the reference image is well corrected. Then, based on the aligned image, the geometric positions of each pixel are recognized, and subsequent detection and analysis processes are performed based on the generated geometric feature map. After the current analysis is completed, the analysis of the next data is carried out.

[0048] That is to say, in this embodiment, first, the initial image can be aligned with a preset reference image to obtain an aligned image, and then the aligned image is used as the image to be recognized. In this way, by aligning the image to be recognized with the preset reference image before image recognition, it can ensure that the data offset caused by the error in the data acquisition process between the image to be processed and the reference image is well corrected, which helps to improve the accuracy of image recognition. And it can be understood that the above preset reference image is an error-free image that is preset and corresponds one-to-one with the initial image before the input image data is obtained. And the above initial image is the image that needs to be recognized in the mask detection process, including any one or several of the first image containing the corresponding vertices of the polygon designed by electronic design automation software, the second image obtained by rasterizing the corresponding vertices of the polygon, and the third image obtained by imaging the mask. That is to say, the input image data obtained in this embodiment can be various types of data. In this way, the input of image data in this application can be various types of images. When the input image is of different types, there is no need to adjust the recognition scheme every time, and different types of input images can be recognized flexibly.

[0049] Correspondingly, when aligning the initial image with the preset reference image, the initial image and the corresponding preset reference image can be obtained, and then the initial image and the preset reference image are preprocessed to obtain a preprocessed initial image and a preprocessed reference image, and the preprocessed initial image is aligned based on the preprocessed reference image; wherein, the above preprocessing includes image denoising and / or image gray level matching. Thus, image alignment is performed between the preprocessed initial image and the reference image to better correct the data offset when obtaining the initial image.

[0050] That is to say, based on the Figure 1 image recognition system shown in Figure 3As shown, the data input by the user includes the image to be recognized. Then, the detection control system can determine the initially created initial classification tree corresponding to the image category of the image to be recognized based on the type of the image to be recognized. That is, before determining the initially created initial classification tree corresponding to the image category of the image to be recognized, the image categories to be recognized can be determined first, and the feature categories of the images of each image category can be counted to obtain the set of feature categories corresponding to each image category respectively. Then, based on the above image categories and the set of feature categories, the initial classification trees corresponding to each image category are created; wherein, each level of the above initial classification tree is completely filled, and each leaf node of the initial classification tree corresponds to each feature category in the corresponding set of feature categories. It can be understood that, in a specific embodiment, the above classification tree can be a binary classification tree. At this time, the initially created initial classification tree is the corresponding maximum tree. However, the classification tree adopted in this embodiment is not limited to the binary classification tree, and any classification tree that can achieve the corresponding classification effect can implement the above solution. In this way, by pre-constructing the corresponding maximum classification tree as the initial classification tree based on the feature categories corresponding to each image category, during subsequent image recognition, the structure of the classification tree can be adjusted according to different target features to be recognized to obtain the target classification tree, so as to improve the speed and flexibility of image feature classification.

[0051] Step S12: Adjust the structure of the initial classification tree to obtain a target classification tree; each leaf node in the target classification tree corresponds to each target feature category to be recognized in the current image to be recognized.

[0052] In this embodiment, as Figure 4 shown, after obtaining the image data to be processed, the process of generating the corresponding classification tree can be carried out based on the input data. Specifically, in this embodiment, it is first necessary to generate a multi-level classification tree MultiLayer_Tree based on the input of the algorithm parameter Parameters_Input. That is to say, in this embodiment, after obtaining the file that defines the overall pattern type and the geometric position categories required for recognition of each pattern (the algorithm parameter can include the pattern type in the current layer / mask of the current data, etc.) and the file that defines the overall maximum multi-level classification tree, the file input can be processed to output the initialized multi-level classification tree, that is, the target classification tree, and each leaf node in the target classification tree corresponds to each target feature category to be recognized in the current image to be recognized.

[0053] Based on the above technical solution, it can be understood that in the system maintenance process of this embodiment, for different types of mask layer data input, a file will be pre-stored locally. The data in the file is stored in a tree-like data structure. The tree in this file is the largest complete tree, and this tree structure is generally a multi-layer tree structure. Each layer represents a classification operation / geometric feature refinement. Through multiple classification operations and feature refinements, leaf nodes are finally obtained, and each leaf node corresponds to a geometric position category. Then, based on the above file, a multi-level classification tree can be generated according to the input file and parameters. That is to say, in this embodiment, the initially input tree is the complete tree of the geometric position categories corresponding to the current pattern type. Through the pattern type file and the file defining the overall pattern type and the geometric position categories required for each pattern, nodes of the complete tree can be removed, and only the required nodes are retained. By doing so, the recognition efficiency can be improved by reducing the nodes, that is, reducing the operations. At the same time, only the leaf nodes corresponding to the geometric position categories to be recognized are retained. That is to say, as Figure 4 shown, when performing recognition, each initialized tree used for recognition and classification in subsequent steps is a subtree of the complete tree.

[0054] In a specific embodiment, after selecting the input of the pattern type to be classified in this embodiment, through multi-level judgments of one or more attributes, the levels and nodes of the classification tree are initialized. Each node can have more than two child nodes according to the subsequent classification input attributes. The number of final leaf nodes corresponds to the number of categories to be classified, and each category to be classified corresponds to a leaf node. As Figure 5 shown in a generated multi-level classification tree structure, when the current wiring image, that is, the image to be recognized, is input, the corresponding complete tree can be determined according to the type of the pattern to be classified, and then the subsequent nodes can be determined. As Figure 5 the multi-level classification judgment process in can include the near-edge area (NearEdge) and the flat area (Flat Area). The near-edge area usually refers to the area near the mask pattern edge affected by optical or process effects, emphasizing the sensitive area of the mask pattern edge. The flat area is the area on the mask where no patterns are designed, where the alignment marks between the mask and the wafer can be placed, or it can be used as a reference area for process testing or defect detection. As Figure 5As shown, during the process of geometric feature classification, feature classification is mainly performed on the near-edge region, and generally no geometric feature classification process is required for the region without graphics. Specifically, the above-mentioned near-edge region can further include an edge (Edge) and a corner (Corner). The edge (Edge) can be further divided into geometric feature 1 (Geo-Feature1) and geometric feature 2 (Geo-Feature2). Among them, Geo-Feature refers to the key attributes such as the geometric shape, size, position, and topological structure of the mask pattern. These features directly affect the accuracy of pattern transfer to the wafer in the lithography process and are the core analysis objects for optical proximity correction (OPC), process simulation, and yield optimization.

[0055] And it should be noted that, based on the above technical solution, as Figure 3 shown, before the start of the next image recognition, in this embodiment, it is only necessary to determine the image type of the image to be recognized currently and obtain the target features that need to be recognized for this image, and then the leaf node corresponding to the target feature can be determined according to the complete tree corresponding to the image type, so as to determine the subtree corresponding to this image. Based on this subtree, the corresponding image recognition process can be executed, thereby indicating that corresponding recognition schemes are re-established for different types of images, effectively improving the efficiency of recognizing features in the mask image. In this way, by pre-constructing the corresponding maximum classification tree for each feature category corresponding to each image category as the initial classification tree, during subsequent image recognition, the structure of the classification tree can be adjusted according to different target features to be recognized to obtain the target classification tree, so as to improve the speed and flexibility of image feature classification.

[0056] Step S13: Identify the target image features of the image to be recognized, and determine the corresponding image recognition result based on the target image features and the target feature categories corresponding to the leaf nodes in the target classification tree.

[0057] In this embodiment, based on the above target classification tree, the target image features of the image to be recognized can be identified, and the corresponding image recognition result can be determined based on the target image features and the target feature categories corresponding to the leaf nodes in the target classification tree. As Figure 3As shown, after the geometric position of the input image is recognized based on the target classification tree, the corresponding geometric feature map FeatureMap can be generated based on the recognized feature group Feature_Group. That is to say, in this embodiment, the overall structure of the classification of the data Data_Input is initialized through tree-shaped classification to obtain the multi-level classification tree MultiLayer_Tree, and feature recognition is sequentially performed based on MultiLayer_Tree and Feature_Group to obtain FeatureMap. In the overall feature classification architecture, this embodiment adopts a multi-level tree structure for classification. In this way, since the number of layers and the structure of the multi-level tree can be flexibly changed, a more flexible feature recognition effect can be achieved for different numbers of categories and types of input requirements.

[0058] Based on the above technical solution, in this embodiment, the initial classification tree pre-created corresponding to the image category of the image to be recognized is first determined, and the structure of the initial classification tree is adjusted to obtain the target classification tree, where each leaf node in the target classification tree corresponds to each target feature category to be recognized in the current image to be recognized. Then, the target image features of the image to be recognized are recognized, and based on the target image features and the target feature categories corresponding to each leaf node in the target classification tree, the corresponding image recognition result is determined. In this way, classification can be performed through a multi-level tree structure. In this way, since the number of layers and the structure of the multi-level tree can be flexibly changed, for different numbers, different types of feature categories to be recognized, and different image input requirements, this embodiment solves the problem of complex recognition processes caused by the increasing complexity of current integrated circuit patterns and the increasing number of geometric position categories on the patterns, and can flexibly change the classification method according to the types and numbers of geometric position categories to be recognized, achieving a more flexible recognition effect.

[0059] Embodiment 2

[0060] Based on the previous embodiment, it can be seen that this application can perform classification through a multi-level tree structure to achieve feature recognition of different numbers, different types of features, and different image inputs. Next, in this embodiment, the process of recognizing the image features in the image to be recognized will be described in detail. Refer to Figure 6 As shown, this embodiment provides a specific feature recognition method, including:

[0061] Step S21: Perform binarization processing on the image to be recognized to obtain a binarized image, and process the binarized image to obtain the target image features of the image to be recognized.

[0062] In this embodiment, as Figure 4After obtaining the image to be recognized as shown, the corresponding preprocessing process and the edge recognition of the mask pattern can be performed on the image to be recognized. Specifically, when the detection control system obtains the input data to be processed Data_Input and the algorithm parameters Parameters_Input, it preprocesses the input data Data_Input and performs edge recognition of the mask pattern to obtain the preprocessed data PreProcessed_Data and the edge data Edge_Data. That is to say, first, the data pattern to be processed (i.e., the image to be recognized) and the algorithm parameters can be input. Among them, the above algorithm parameters are all parameters available for geometric category-related recognition, including the size of each geometric feature neighborhood, the recognition sensitivity of each geometric feature, the priority of each category, and so on. After that, a preset processing process including preprocessing and edge recognition is performed to obtain the preprocessed data and the recognized edge data.

[0063] Specifically, when processing a binary image to obtain target image features, first, edge recognition can be performed on the binary image to obtain a target edge image, and defective edges in the binary image and the target edge image can be recognized. Then, based on the position information corresponding to the defective edges, the first target image features of the image to be recognized can be determined. Correspondingly, when performing edge recognition on the binary image to obtain a target edge image, edge recognition can be first performed on the binary image to obtain an initial edge image, and a connectivity test can be performed on the edge patterns in the initial edge image. Then, based on the test results, the target edge lines in the edge patterns are connected to obtain the target edge image; where the above target edge lines are discontinuous edge lines in the edge patterns. That is to say, in this embodiment, different numbers of categories, different types of geometric position types, and various algorithm parameters corresponding to the types can be input. At the same time, different types of data to be processed, Data_Input, are input, that is, Data_Input can be various different types of data. For example, the vertices corresponding to polygons in a pattern designed by electronic design automation software, the rasterized image obtained after rasterizing the above vertices, or the actual captured image obtained after imaging a mask using an optical imaging element can be input. Next, preprocessing such as noise reduction and image binarization will be first performed on the input data to be processed, and then pattern edge recognition will be performed on the data. In this process, it can be understood that due to the influence of image resolution, problems such as discontinuous edges may exist. Therefore, in this embodiment, a connectivity test and connection need to be performed on the discontinuous edge situation to enter the next step of processing. In this way, before performing edge recognition on the image, image binarization processing is first performed to facilitate more convenient and efficient extraction of image features in the subsequent process. And during the process of performing edge recognition on the binarized image to obtain the edge defects therein, the edge data Edge_Data is geometrically classified by using the edge defect detection method instead of the traditional method, and a feature group Feature_Group is obtained. In this way, in this embodiment, the edge defect detection method is used instead of the traditional method (template matching, feature vector-based recognition, or the method of judging line width after pattern thinning, etc.) for geometric classification. Since the edge defect has higher accuracy and higher recognition sensitivity, the recognition accuracy of the data after optical proximity correction and sub-resolution assist feature processing can be improved; and since the detection of edge defects has no specific requirements for the input of data, when the types of input image data are different, the efficiency can be further improved. It can be understood that the edge defects described in this embodiment are edge perturbation terms recognized on the edge, and their length, depth, and area are within a certain threshold range. At the same time, in this embodiment, considering that when performing pattern edge recognition on the data, due to the influence of image resolution, problems such as discontinuous edges may exist, a connectivity test and connection can be performed on the discontinuous edge situation to improve the effect of image edge recognition.

[0064] As shown Figure 4 in the figure, in order to further improve the accuracy during feature recognition, corresponding edge smoothing, edge defect detection, feature quantization, and minimum rectangle segmentation processing can also be performed in this embodiment to obtain target image features with higher accuracy. That is to say, in this embodiment, based on the preprocessed data PreProcessed_Data (i.e., the binary image) and the edge data Edge_Data (i.e., the target edge image), first perform edge smoothing on Edge_Data to obtain the smoothed Smoothed_Edge, and perform edge defect detection and feature quantization on PreProcessed_Data and Smoothed_Edge, etc., to obtain the feature group Feature_Group. Correspondingly, based on the above embodiment, that is, input the preprocessed (binary) data and the recognized edge data, and then perform edge defect detection and feature quantization to output the feature group Feature_Group. It should be noted that the above-mentioned edge defects are used to characterize the one-dimensional fluctuations within a certain quantization range recognized on the edge.

[0065] That is to say, in this embodiment, when recognizing the defective edges in the binary image and the target edge image, the target edge image can be first smoothed to obtain the target edge image after edge smoothing, and the defective edge recognition is performed on the binary image and the target edge image after edge smoothing. Currently, in order to overcome the influence of a certain degree of optical diffraction limit, resolution enhancement techniques such as optical proximity correction and sub-resolution assist features are introduced in the mask design process. However, optical proximity correction may cause perturbations to the size and edges of printable features. At the same time, sub-resolution assist features make the pattern of the mask more complex by adding some non-printable features, resulting in possible pattern perturbations during the printing process. Therefore, considering the problem that the pattern edges become very complex after adding various features, before using edge features for geometric recognition in this embodiment, the image edges can be smoothed to avoid misjudgment caused by the relatively complex pattern edges of the mask due to the addition of techniques such as sub-resolution assist features. That is to say, by first performing edge smoothing on the data Edge_Data to obtain Smoothed_Edge and using it as one of the inputs for defect detection and feature quantization, an edge smoothing step can be added before using edge features for geometric recognition, which can reduce the possibility of misjudgment in the geometric recognition process caused by assist features such as sub-resolution assist features used to improve the optical diffraction effect.

[0066] In a specific embodiment, in this embodiment, it is possible to detect each pattern edge with protrusions or depressions in the binarized image and the target edge image, and determine the defective edge from each pattern edge based on the geometric parameters of the pattern edge; wherein, the above-mentioned defective edge is a pattern edge whose geometric parameters meet the preset defective edge condition. In this way, through the detection and quantification of edge defects, the recognition accuracy of the data after optical proximity correction and sub-resolution assist feature processing is greatly improved, thereby improving the robustness of the geometric classification result. It can be understood that when smoothing the edge in this embodiment, optional smoothing techniques include polygon approximation, morphological processing, contour downsampling, Fourier transform and frequency domain truncation, etc., which can be selected according to the actual processing requirements of the mask image. And it should be noted that in the smoothing process, the degree of contour smoothing can be selected through parameters or sliding windows to determine whether to smooth out all sub-resolution assist features, etc. As Figure 7 shown in the examples of different degrees of contour smoothing, Figure 7 a is the original image before smoothing processing, Figure 7 b is the weak smoothing mode, Figure 7 c is the strong smoothing mode. It can be seen that the weak smoothing mode retains all sub-pixel assist features, while the strong smoothing mode eliminates all sub-pixel assist features and makes the edges of the pattern smoother at the same time, which is helpful for the detection of stronger edge defects, but not conducive to the detection of weaker edge defects. Therefore, in this embodiment, different smoothing modes can effectively avoid the perturbation of the size and edge of printable features caused by optical proximity correction, and the pattern perturbation caused by sub-resolution assist features by adding non-printable features.

[0067] And in this embodiment, after edge smoothing, it is necessary to define the concept of edge defect. It should be noted that the edge defect in this embodiment is an edge perturbation term whose length, depth, and area are within a certain threshold range and is recognized on the edge. For example, three judgment criteria are set: area, width, and height, and corresponding thresholds are preset. When judging edge defects, it can be set according to the actual situation that when one or more of the above judgment criteria meet the corresponding threshold conditions, it is determined as an edge defect. For example Figure 8 in, it can be set that when the area, width, and height of a certain edge perturbation all meet the preset threshold, it is determined as an edge defect. As Figure 8 in, the edge perturbation on the left is determined as an edge defect because its area, length, and depth are all within a certain threshold range, that is, AREA Def_lthresh <AREA<AREA Def_hthresh 、H Def_lthresh <H<H Def_hthresh 、W Def_lthresh <W<W Def_hthresh, so it is considered as an edge defect. For the protrusion on the right edge, since the quantization value of its edge defect is not within the set threshold range, i.e., AREA < AREA Def_lthresh , H < H Def_lthresh , W < W Def_lthresh , so it is considered only a small edge perturbation and is not recognized as an edge defect. The quantization value of the edge defect here is not limited to the area, length, and depth described above. Any value that can be used to describe the protrusion, depression, or perturbation amount on the edge can be considered as the quantization standard for the edge defect. In this way, through the unified identification, detection, and quantization of edge defects, one-dimensional edges can be quantitatively described, so as to more accurately describe the geometric positions of the pixels near the edge and at the edge.

[0068] And in this embodiment, the dimensionality reduction processing can also be performed on the pattern edge based on the above geometric parameters to obtain one-dimensional features for describing the pattern edge, and the above one-dimensional features can be used to determine the defective edge from each pattern edge. That is to say, in this embodiment, dimensionality reduction processing can be performed on the edge, and one-dimensional edge data can be obtained through contour description. In a specific embodiment, the horizontal axis of the data corresponds to the position of each edge point, and the vertical axis corresponds to the features used to describe the edge. For example, the overall center of gravity of the pattern can be obtained, and the distances from each point of a connected edge to the center of gravity can be described as one-dimensional description data features. This feature can be used as an auxiliary in the edge defect quantization process to obtain a more accurate edge defect position distribution. In this way, through the dimensionality reduction processing of the pattern edge, the changes of one-dimensional edges can be quantitatively described in detail, and the recognition accuracy of the data after optical proximity correction and sub-resolution assist feature processing is greatly improved. In this way, by performing dimensionality reduction on the data Smoothed_Edge, the one-dimensional curve OneDim_Curve after dimensionality reduction is obtained, and at the same time, its result is combined with other feature groups Feature_Group obtained based on the image data PreProcessed_Data to obtain an updated feature group Feature_Group, which is used as the basis for the feature recognition process. The one-dimensional curve obtained by dimensionality reduction can be combined with two-dimensional image processing. Since both one-dimensional and two-dimensional data features are considered, a more reliable edge defect quantization result can be obtained.

[0069] In another specific embodiment, in the process of processing the binary image to obtain the target image features of the image to be recognized, the binary image can also be line-scanned based on a target number of preset line-scanning directions, and for the binary image regions scanned in each preset line-scanning direction, corresponding minimum segmentation rectangles are generated, so as to determine the second target image features of the image to be recognized according to each minimum segmentation rectangle. Specifically, the minimum rectangle segmentation can be performed based on the data PreProcessed_Data to expand the obtained feature group Feature_Group. That is to say, in this embodiment, through the preprocessed data and the feature group Feature_Group obtained above, by performing minimum rectangle scanning and segmentation, the expanded feature group Feature_Group can be output. Specifically, as follows Figure 9 shown, for the data after edge smoothing, in this embodiment, line scanning in a set direction can be performed to obtain the minimum rectangles in each direction, Figure 9 a is the binary image, Figure 9 b is the minimum rectangle segmentation result when performing 0° horizontal scanning. In this way, by performing minimum rectangle scanning and segmentation on the data, adding new feature categories, the pattern structure can be refined, which helps to improve the judgment accuracy of subsequent recognition processes. And by further determining the features within and between the minimum rectangles, the feature group can be expanded, further improving the robustness of the geometric classification result. In this embodiment, for the smoothed data, it can be selected to perform minimum rectangle segmentation on the data PreProcessed_Data, judge the relationships within and between the patterns according to the relationships between the rectangles, and expand the feature group Feature_Group based on the result as the basis for the feature recognition process. In this way, according to the combination of different feature quantifications such as edge defect quantification and within and between minimum rectangles, more information can be recognized by expanding the feature group, which ensures a certain degree of robustness for the geometric classification result.

[0070] Step S22: Determine the corresponding image recognition result based on the target image features and the target feature categories corresponding to the leaf nodes in the target classification tree.

[0071] In this embodiment, based on the target image features and the target feature categories corresponding to the leaf nodes in the target classification tree, the corresponding image recognition result can be determined. It can be understood that the above-mentioned target image features include the first target image feature and the second target image feature, and the first target image feature and the second target image feature can be used as the basis for feature classification respectively. The above-mentioned first target image feature is the feature combined with the edge-smoothed feature and the one-dimensional feature after dimensionality reduction, and the second target image feature is the feature obtained by line scanning. Specifically, in this embodiment, each target image feature can be matched with each target feature category to obtain the corresponding matching result, and a geometric feature map of the image to be recognized can be generated according to the matching result. And determine the priority between the target feature categories corresponding to each leaf node, and generate a geometric feature map of the image to be recognized based on the priority and the matching result. That is to say, based on MultiLayer_Tree and Feature_Group, this embodiment can perform sequential recognition of the leaf nodes of the classification tree and output the final result feature map FeatureMap. That is, obtain the multi-level classification tree, the total feature group Feature_Group, and the priority of each geometric category obtained in the above embodiment, and by sequentially operating and recognizing each node on each tree, the corresponding feature map can be output. It can be understood that in this embodiment, the above-mentioned total feature group is a feature group composed of the edge-smoothed feature, the one-dimensional feature after dimensionality reduction, and the feature obtained by line scanning.

[0072] In this embodiment, it should be noted that, such as Figure 10As shown, the geometric position recognition in this embodiment is a pixel-level classification process used to describe the geometric type of the neighborhood where each pixel in the image is located, such as isolated edge, edge-to-edge, isolated corner, and so on. In a specific embodiment, single corners and inner corners in the mask image can be recognized. It can be understood that the above-mentioned single corner refers to an independent single corner in the mask pattern that is not directly affected by other adjacent structures, usually located at the periphery or isolated area of the pattern. For example, the end corners of a rectangular line (such as the two corner points on both sides of the line end) and the tip corners of isolated raised structures (such as contact holes or bumps). Its characteristics are: there is no adjacent side extension, and it is only formed by the intersection of two sides. Therefore, it is vulnerable to optical diffraction effects in lithography, resulting in corner rounding. The above-mentioned inner corner refers to the included angle formed by two sides indented inward in the mask pattern, commonly found inside grooves, trenches, or complex graphics. For example, the bottom corner of a U-shaped groove and the internal concave corner formed by the intersection of dense lines (such as the inner corner of a "T"-shaped intersection structure). Its characteristics are: the included angle is usually less than 180° (such as a 90° right angle or a sharper angle). Therefore, the space is narrow, and problems such as incomplete material filling or etching are likely to occur in the process. Based on the above characteristics, this embodiment can recognize the above geometric position features to better determine the edge defects at the corresponding positions and avoid greater losses in the future. The above process takes various data related to the mask pattern as input. After uniformly converting the various data into an image Data_Input with dimensions of m*n, through geometric position recognition processing, a result feature map FeatureMap with the same dimensions of m*n is output. The gray value of each pixel on this FeatureMap represents the geometric feature of the neighborhood where the pixel is located. Take Figure 10 as an example. When an image is input, assuming there are three categories in the image: single corner, inner corner, and others. After the geometric position recognition process, a FeatureMap is output. The value of each pixel on this FeatureMap corresponds to the three categories respectively (assuming the value of single corner is 0, the value of inner corner is 1, and the value of others is 2). During the recognition process, if there are some pixels that belong to two or more geometric position categories at the same time, the value of these pixels on the FeatureMap will be based on the priority between the categories. Assuming the priority here is single corner > inner corner > others, the pixel values in the overlapping area of the single corner and inner corner features on the FeatureMap will be the value corresponding to the single corner, that is, the pixel value here is 0.

[0073] In this way, based on the above-mentioned target image features, through the relationships such as the edge defect quantization result, the long-side distance between each rectangle, connectivity, the number and length of overlapping edges between each rectangle, and the distance between rectangles, the determination of each node of the target classification tree can be carried out respectively, so as to obtain the classification results of specific categories for each leaf node of the final classification tree, where each leaf node corresponds to a specific geometric position classification category. At this time, each obtained leaf node corresponds to a geometric map GeoMap of a geometric feature. Each GeoMap is a binary map, where 0 in the map indicates not being this geometric feature, and 1 indicates being this geometric feature type. After obtaining the GeoMaps of the number of geometric features to be classified, according to the priority of each feature, from high to low priority, the pixel value of the pixel at the same image position on the GeoMap of each specific feature is used to determine the pixel value of the pixel at this image position on the final FeatureMap, so as to distribute each feature on the final FeatureMap respectively, and the final FeatureMap is obtained. In this way, by performing geometric position classification on the target classification tree obtained by excising nodes from the complete tree, the recognition efficiency can be improved by reducing nodes, that is, reducing operations, and generating several feature maps based on the priority between the target feature categories corresponding to different leaf nodes can also effectively improve the practicability of the geometric feature map.

[0074] Based on the above technical solution, in this embodiment, after inputting the data to be processed Data_Input and the algorithm parameters Parameters_Input, that is, based on Parameters_Input, a multi-level classification tree MultiLayer_Tree is generated, and the input data Data_Input is preprocessed and the edge of the mask pattern is recognized to obtain the preprocessed data PreProcessed_Data and the edge data Edge_Data. Then, based on the data PreProcessed_Data and Edge_Data, smoothed edge defect detection and quantization are performed to obtain the feature group Feature_Group, and based on the data PreProcessed_Data, minimum rectangle segmentation is performed to expand the feature group Feature_Group. Finally, based on MultiLayer_Tree and Feature_Group, the leaf nodes of the classification tree are sequentially recognized, and the final result feature map FeatureMap is output. As Figure 11 shown, when inputting specific parameters, it may specifically include but is not limited to: parameters required for methods / algorithms adopted in this embodiment such as edge recognition sensitivity, OPC feature recognition sensitivity, and SRAFs feature recognition sensitivity, which can be adjusted accordingly according to actual needs. Based on Figure 1In the feature recognition process of the system shown, all the above-mentioned contents are implemented using program code in the storage. First, a design tool can be used to design pattern data with specific geometric patterns. After inputting the pattern through the interface, the number and types of categories can be flexibly changed. Then, relevant parameters such as edge smoothness and sub-pixel resolution-assisted feature removal are set through the host computer interface. And in this process, relevant parameters such as edge defect detection-related parameters and dimensionality reduction-related parameters can exist or have a large number of options for corresponding adjustment. After the parameters are confirmed, multiple specific parameters are input into the algorithm, and geometric position recognition based on specific parameters is performed respectively. Then, a feature map corresponding to the required number of classifications is output, and the data to be detected is loaded onto the user system during the inference process. In this way, the geometric position recognition in this embodiment is based on the form of a multi-level classification tree. At the same time, according to the edge defect quantization result of the mask pattern and the relationship between rectangles after minimum rectangle segmentation, the leaf nodes of each level of the classification tree are recognized. Each leaf node corresponds to a geometric position feature. After obtaining the result, the result is output. And by using a brand-new concept of edge defect during the recognition process, the detection and quantization of edge defects can be realized, and the change of one-dimensional edges can be described in a refined quantitative manner, greatly improving the recognition accuracy of the data after optical proximity correction and sub-resolution assisted feature processing. At the same time, various fluctuations on the edge can be quantified using data, thus ensuring a certain degree of robustness of the geometric classification result.

[0075] Embodiment III

[0076] Based on the above embodiments, refer to Figure 12 As shown, this embodiment also provides an image recognition device, including:

[0077] A classification tree determination module 11, configured to determine an initial classification tree pre-created corresponding to the image category of the image to be recognized;

[0078] A classification tree adjustment module 12, configured to perform structural adjustment on the initial classification tree to obtain a target classification tree; each leaf node in the target classification tree corresponds to each target feature category to be recognized in the current image to be recognized;

[0079] A feature recognition module 13, configured to recognize the target image features of the image to be recognized, and determine the corresponding image recognition result based on the target image features and the target feature categories corresponding to the leaf nodes in the target classification tree.

[0080] In this embodiment, an initial classification tree corresponding to the image category of the image to be recognized is first determined, and the structure of the initial classification tree is adjusted to obtain a target classification tree, where each leaf node in the target classification tree corresponds to each target feature category to be recognized in the current image to be recognized. Then, the target image features of the image to be recognized are recognized, and based on the target image features and the target feature categories corresponding to the leaf nodes in the target classification tree, the corresponding image recognition result is determined. Through the above technical solution, classification can be performed through a multi-level tree structure. In this way, since the number of layers and the structure of the multi-level tree can be flexibly changed, a more flexible recognition effect is achieved for different numbers, different types of feature categories to be recognized, and different image input requirements.

[0081] In some specific embodiments, the image recognition device further includes:

[0082] An image alignment module, configured to perform image alignment on the initial image based on a preset reference image to obtain an aligned image; and use the aligned image as the image to be recognized.

[0083] The initial image is an image to be recognized in the mask detection process, and includes any one or more of a first image containing corresponding vertices of a polygon designed by electronic design automation software, a second image obtained by rasterizing the corresponding vertices of the polygon, and a third image obtained by imaging the mask.

[0084] In some specific embodiments, the image alignment module specifically includes:

[0085] An image acquisition unit, configured to acquire an initial image and a corresponding preset reference image;

[0086] An image preprocessing unit, configured to preprocess the initial image and the preset reference image to obtain a preprocessed initial image and a preprocessed reference image;

[0087] An image alignment unit, configured to align the preprocessed initial image based on the preprocessed reference image;

[0088] Wherein, the preprocessing includes image noise reduction and / or image gray level matching.

[0089] In some specific embodiments, the image recognition device further includes:

[0090] An image determination module, configured to determine each image category;

[0091] A category determination module, configured to count each feature category possessed by the images of each image category to obtain a set of feature categories corresponding to each image category respectively;

[0092] A classification tree creation module, configured to create respective initial classification trees corresponding to each of the image categories based on the image categories and the set of feature categories;

[0093] Wherein, each level of the initial classification tree is completely filled, and each leaf node of the initial classification tree corresponds to each feature category in the corresponding set of feature categories.

[0094] In some specific embodiments, the feature recognition module 13 specifically includes:

[0095] An image binarization sub-module, configured to perform binarization processing on the image to be recognized to obtain a binarized image;

[0096] An image processing sub-module, configured to process the binarized image to obtain the target image features of the image to be recognized.

[0097] In some specific embodiments, the image processing sub-module specifically includes:

[0098] An edge recognition unit, configured to perform edge recognition on the binarized image to obtain a target edge image;

[0099] A defect recognition unit, configured to recognize defect edges in the binarized image and the target edge image;

[0100] A feature determination unit, configured to determine the first target image features of the image to be recognized based on the position information corresponding to the defect edges.

[0101] In some specific embodiments, the edge recognition unit is specifically configured to:

[0102] Perform edge recognition on the binarized image to obtain an initial edge image;

[0103] Perform a connectivity test on the edge patterns in the initial edge image, and connect the target edge lines in the edge patterns based on the test results to obtain a target edge image;

[0104] Wherein, the target edge lines are discontinuous edge lines in the edge patterns.

[0105] In some specific embodiments, the defect recognition unit is specifically configured to:

[0106] Perform edge smoothing on the target edge image to obtain the target edge image after edge smoothing;

[0107] Perform defect edge recognition on the binarized image and the target edge image after edge smoothing.

[0108] In some specific embodiments, the defect recognition unit is specifically configured to:

[0109] Detect each pattern edge with protrusions or depressions in the binary image and the target edge image;

[0110] Based on the geometric parameters of the pattern edges, determine the defect edges from each of the pattern edges;

[0111] Wherein, the defect edge is a pattern edge whose geometric parameters meet the preset defect edge condition.

[0112] In some specific embodiments, the defect recognition unit is specifically configured to:

[0113] Perform dimensionality reduction processing on the pattern edges based on the geometric parameters to obtain one-dimensional features for describing the pattern edges;

[0114] Use the one-dimensional features to determine the defect edges from each of the pattern edges.

[0115] In some specific embodiments, the image processing sub-module specifically includes:

[0116] An image line scanning unit for line scanning the binary image based on a target number of preset line scanning directions;

[0117] A rectangle generation unit for generating the minimum rectangle results after segmentation in each direction corresponding to the binary image region scanned in each of the preset line scanning directions;

[0118] A feature determination unit for determining the second target image feature of the image to be recognized according to the minimum rectangle results after segmentation in each direction.

[0119] In some specific embodiments, the feature recognition module 13 specifically includes:

[0120] A feature matching sub-module for matching each of the target image features with each of the target feature categories to obtain corresponding matching results;

[0121] A feature map generation sub-module for generating the geometric feature map of the image to be recognized according to the matching results.

[0122] In some specific embodiments, the feature map generation sub-module specifically includes:

[0123] A priority determination unit for determining the priorities among the target feature categories corresponding to each of the leaf nodes;

[0124] A feature map generation unit for generating the geometric feature map of the image to be recognized based on the priorities and the matching results.

[0125] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The program can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above methods. Among them, the storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM), etc.

[0126] Furthermore, the embodiments of the present application also disclose an electronic device. Figure 13 It is a structural diagram of an electronic device 20 shown according to an exemplary embodiment. The content in the figure should not be regarded as any limitation on the scope of use of the present application.

[0127] Figure 13 It is a schematic structural diagram of an electronic device 20 provided by the embodiments of the present application. The electronic device 20 may specifically include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. Among them, the memory 22 is used to store a computer program, and the computer program is loaded and executed by the processor 21 to implement the relevant steps in the image recognition method disclosed in any of the foregoing embodiments. In addition, the electronic device 20 in this embodiment may specifically be an electronic computer.

[0128] In this embodiment, the power supply 23 is used to provide working voltage for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and external devices, and the communication protocol it follows is any communication protocol applicable to the technical solution of the present application, and no specific limitation is imposed on it here; the input / output interface 25 is used to obtain external input data or output data to the outside, and its specific interface type can be selected according to specific application needs, and no specific limitation is made here.

[0129] In addition, as a carrier for resource storage, the memory 22 can be a read-only memory, a random access memory, a magnetic disk, or an optical disk, etc. The resources stored thereon can include an operating system 221, a computer program 222, etc., and the storage method can be temporary storage or permanent storage.

[0130] Among them, the operating system 221 is used to manage and control each hardware device and computer program 222 on the electronic device 20, and it can be Windows Server, Netware, Unix, Linux, etc. In addition to the computer program that can be used to complete the image recognition method executed by the electronic device 20 disclosed in any of the foregoing embodiments, the computer program 222 can further include computer programs that can be used to complete other specific tasks.

[0131] Furthermore, the present application also discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, the foregoing disclosed image recognition method is implemented. For the specific steps of this method, reference can be made to the corresponding content disclosed in the foregoing embodiments, and details will not be elaborated herein.

[0132] Furthermore, the embodiments of the present application also disclose a computer program product, including a computer program / instructions, and when the computer program / instructions are executed by a processor, the steps of the image recognition method disclosed in any of the foregoing embodiments are implemented.

[0133] In this specification, the various embodiments are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The same or similar parts among the various embodiments can be referred to each other. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the description in the method section.

[0134] Those skilled in the art can further realize that the units and algorithm steps of the examples described in combination with the embodiments disclosed in this article can be implemented by electronic hardware, computer software, or a combination of the two. To clearly illustrate the interchangeability of hardware and software, the components and steps of the examples have been generally described according to their functions in the above description. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present application.

[0135] The steps of the method or algorithm described in combination with the embodiments disclosed in this article can be directly implemented by hardware, a software module executed by a processor, or a combination of the two. The software module can be placed in a random access memory (RAM), internal memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, register, hard disk, removable disk, CD-ROM, or any other form of storage medium well-known in the technical field.

[0136] Finally, it should also be noted that in this text, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or device. Without further limitation, an element defined by the statement "comprising an..." does not exclude the presence of additional identical elements in the process, method, article or device comprising the said element.

[0137] The above preferred embodiments have further elaborated on the purpose, technical solutions and advantages of the present application. It should be understood that the above description is only the preferred embodiments of the present application and is not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application shall be included within the protection scope of the present application.

Claims

1. An image recognition method, characterized in that: include: Determining a pre-created initial classification tree corresponding to the image category of the image to be identified; Structural adjustment is performed on the initial classification tree to obtain a target classification tree; Each leaf node in the target classification tree corresponds to each target feature category that needs to be identified in the current image to be identified; Identify the target image features of the image to be identified, and determine the corresponding image recognition result based on the target image features and the target feature categories corresponding to each leaf node in the target classification tree.

2. The image recognition method according to claim 1, characterized in that: Also includes: Performing image alignment on the initial image based on a preset reference image to obtain an aligned image; The aligned image is used as the image to be recognized.

3. The image recognition method according to claim 2, characterized in that: The initial image is an image that needs to be identified during the mask detection process, including any one or more of a first image containing corresponding vertices of a polygon designed using electronic design automation software, a second image obtained by rasterizing the corresponding vertices of the polygon, and a third image obtained by imaging the mask.

4. The image recognition method according to claim 2, characterized in that: The performing image alignment on the initial image based on the preset reference image comprises: Acquire an initial image and a corresponding preset reference image; Preprocessing the initial image and the preset reference image to obtain a preprocessed initial image and a preprocessed reference image; aligning the preprocessed initial image based on the preprocessed reference image; Wherein, the preprocessing includes image denoising and / or image grayscale matching.

5. The image recognition method according to claim 1, characterized in that: Before determining the pre-created initial classification tree corresponding to the image category of the image to be identified, the method further includes: Determine each image category; Counting the feature categories possessed by the images of each image category to obtain feature category sets corresponding to each image category; Based on the image category and the feature category set, creating initial classification trees corresponding to the image categories respectively; Each level of the initial classification tree is completely filled, and each leaf node of the initial classification tree corresponds to each feature category in the corresponding feature category set.

6. The image recognition method according to any one of claims 1 to 5, characterized in that: The identifying target image features of the image to be identified includes: Binarizing the image to be identified to obtain a binary image; The binary image is processed to obtain target image features of the image to be identified.

7. The image recognition method according to claim 6, characterized in that: The processing of the binary image to obtain the target image features of the image to be identified includes: Performing edge recognition on the binary image to obtain a target edge image; Identifying defect edges in the binary image and the target edge image; Based on the position information corresponding to the defect edge, a first target image feature of the image to be identified is determined.

8. The image recognition method according to claim 7, characterized in that: The performing edge recognition on the binary image to obtain a target edge image includes: Performing edge recognition on the binary image to obtain an initial edge image; Performing a connectivity test on the edge pattern in the initial edge image, and connecting the target edge lines in the edge pattern based on the test result to obtain a target edge image; Wherein, the target edge line is a discontinuous edge line in the edge pattern.

9. The image recognition method according to claim 7 or 8, characterized in that: The identifying defective edges in the binary image and the target edge image comprises: Performing edge smoothing on the target edge image to obtain the target edge image after edge smoothing; Defect edge recognition is performed on the binarized image and the target edge image after edge smoothing.

10. The image recognition method according to any one of claims 7 to 9, characterized in that: The identifying defective edges in the binary image and the target edge image comprises: Detecting each pattern edge that is convex or concave in the binary image and the target edge image; Determining defective edges from each of the pattern edges based on geometric parameters of the pattern edges; Wherein, the defect edge is a pattern edge whose geometric parameters satisfy preset defect edge conditions.

11. The image recognition method according to claim 10, characterized in that: The step of determining a defect edge from each of the pattern edges based on the geometric parameters of the pattern edges comprises: Performing dimensionality reduction processing on the pattern edge based on the geometric parameters to obtain a one-dimensional feature for describing the pattern edge; The one-dimensional feature is used to determine the defect edge from each of the pattern edges.

12. The image recognition method according to any one of claims 6 to 11, characterized in that: The processing of the binary image to obtain the target image features of the image to be identified includes: Performing line scanning on the binary image based on a target number of preset line scanning directions; For each of the binary image regions scanned in the preset line scanning direction, generating the corresponding minimum rectangle results after segmentation in each direction; The second target image feature of the image to be identified is determined according to the minimum rectangle results after segmentation in each direction.

13. The image recognition method according to any one of claims 1 to 12, characterized in that: The determining of the corresponding image recognition result based on the target image feature and the target feature category corresponding to each leaf node in the target classification tree includes: Matching each of the target image features with each of the target feature categories to obtain a corresponding matching result; A geometric feature map of the image to be identified is generated according to the matching result.

14. The image recognition method according to claim 13, characterized in that: Generating a geometric feature map of the image to be identified according to the matching result includes: Determine the priority between the target feature categories corresponding to each of the leaf nodes; A geometric feature map of the image to be identified is generated based on the priority and the matching result.

15. An image recognition device, characterized in that: include: A classification tree determination module, used to determine a pre-created initial classification tree corresponding to the image category of the image to be identified; A classification tree adjustment module, used for adjusting the structure of the initial classification tree to obtain a target classification tree; Each leaf node in the target classification tree corresponds to each target feature category that needs to be identified in the current image to be identified; The feature recognition module is used to recognize the target image features of the image to be recognized, and determine the corresponding image recognition result based on the target image features and the target feature categories corresponding to each leaf node in the target classification tree.

16. An electronic device, characterized in that: The electronic device comprises a processor and a memory; wherein the memory is used to store a computer program, and the computer program is loaded and executed by the processor to implement the image recognition method according to any one of claims 1 to 14.

17. A computer-readable storage medium, characterized in that: Used to store a computer program, which, when executed by a processor, implements the image recognition method according to any one of claims 1 to 14.

18. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the image recognition method according to any one of claims 1 to 14 is implemented.

Images