Image screening method and device, computer equipment and storage medium
By adopting multi-feature dimension feature extraction and similarity matching methods in the field of image recognition, the problem of low image screening accuracy in traditional technology is solved, and more efficient image screening and hardware resource utilization is achieved.
Patent Information
- Application Number
- CN202411492379.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-24
- Publication Date
- 2025-02-14
AI Technical Summary
The prior art has the problem of low accuracy in image screening in the field of image recognition, especially when faced with massive redundant data, the feature extraction of traditional algorithms is fixed, making it difficult to effectively screen.
Provide an image filtering method, by determining the target image and the non-target image in a plurality of images to be screened, feature extraction and similarity matching are performed based on multiple feature extraction dimensions, and the image similarity between the non-target image and the target image is counted, and Retain images with similarity less than or equal to the threshold.
The accuracy of image screening is improved, and through multi-dimensional feature extraction and similarity matching, the overall difference in image features is more accurately reflected, redundant data is reduced, and hardware utilization efficiency is improved.
Smart Images
Figure CN119445156A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of image recognition technology, and in particular to an image screening method, apparatus, computer equipment, storage medium and computer program product. Background Art
[0002] In the industrial field, deep neural network models have been deeply applied to tasks such as image recognition and workpiece detection. However, in actual production work, the models often face a lot of redundant data, which not only takes up a lot of engineers' time for data screening, but also takes up a lot of hardware. Therefore, in the field of image recognition technology, it is necessary to streamline the massive redundant data.
[0003] In the traditional way, traditional algorithms are used, such as grayscale detection, gradient detection and other detection methods. The extraction of these initial screening algorithms is very fixed. Often when we select the corresponding algorithm, the features that need to be processed have been determined, resulting in the problem of low image screening accuracy. Summary of the invention
[0004] Based on this, it is necessary to provide an image screening method, device, computer equipment, computer-readable storage medium and computer program product that can improve the accuracy of image screening in response to the above technical problems.
[0005] In a first aspect, the present application provides an image screening method. The method comprises:
[0006] Determine a target image and at least two non-target images other than the target image from among multiple images to be screened; for each of the images to be screened, perform feature extraction on the images to be screened based on multiple feature extraction dimensions to obtain image features of the images to be screened corresponding to each of the feature extraction dimensions; based on each of the image features, perform feature similarity matching on the target image and each of the non-target images to obtain a similarity set between the target image and each of the non-target images; the similarity set includes feature similarities corresponding to each of the feature extraction dimensions; for each of the non-target images, perform statistics on multiple feature similarities corresponding to the non-target image to determine image similarity between the non-target image and the target image; retain the target image and non-target images whose image similarity is less than or equal to a similarity threshold.
[0007] In one of the embodiments, determining a target image and at least two non-target images other than the target image from a plurality of images to be screened comprises: acquiring a plurality of images to be screened; determining an image quality score of each of the images to be screened; determining a target image from a plurality of images to be screened whose image quality scores satisfy a quality condition; determining a plurality of candidate images from a plurality of images to be screened whose image quality scores do not satisfy the quality condition; and screening at least two non-target images whose image acquisition objects match the target image from each of the candidate images.
[0008] In one of the embodiments, based on multiple feature extraction dimensions, feature extraction is performed on the image to be screened respectively to obtain image features of the image to be screened corresponding to each of the feature extraction dimensions, including: performing image block processing on the image to be screened to obtain multiple block images; for each feature extraction dimension, determining feature extraction parameters matching the feature extraction dimension; performing feature extraction on each block image according to the feature extraction parameters to obtain block image features of each block image; and fusing the features of each block image to obtain image features of the image to be screened corresponding to each of the feature extraction dimensions.
[0009] In one of the embodiments, the statistical analysis of the multiple feature similarities corresponding to the non-target image to determine the image similarity between the non-target image and the target image includes: determining the feature weights configured for each of the feature extraction dimensions; and statistical analysis of the multiple feature similarities corresponding to the non-target image according to each of the feature weights to determine the image similarity between the non-target image and the target image.
[0010] In one of the embodiments, the method further includes: determining a new target image and at least two new non-target images other than the new target image from among the retained multiple non-target images; determining, for each of the new non-target images, a new image similarity between the new non-target image and the new target image; and retaining the new target image and the new non-target images whose new image similarity is less than or equal to the similarity threshold.
[0011] In one of the embodiments, the method further includes: constructing positive samples based on the target image and non-target images whose image similarity is greater than the similarity threshold; constructing negative samples based on non-target images whose image similarity is less than or equal to the similarity threshold; the positive samples and the negative samples are used to train an object recognition model; the object recognition model is used to identify the image acquisition object of the target image.
[0012] In a second aspect, the present application also provides an image screening device. The device comprises:
[0013] An image determination module is used to determine a target image and at least two non-target images other than the target image from a plurality of images to be screened; a feature extraction module is used to perform feature extraction on each of the images to be screened based on a plurality of feature extraction dimensions to obtain image features of the images to be screened corresponding to each of the feature extraction dimensions; a feature matching module is used to perform feature similarity matching on the target image and each of the non-target images based on each of the image features to obtain a similarity set between the target image and each of the non-target images; the similarity set includes feature similarities corresponding to each of the feature extraction dimensions; a statistical module is used to perform statistics on a plurality of feature similarities corresponding to each of the non-target images to determine the image similarity between the non-target image and the target image; a retention module is used to retain the target image and non-target images whose image similarity is less than or equal to a similarity threshold.
[0014] In a third aspect, the present application further provides a computer device. The computer device includes a memory and a processor, the memory stores a computer program, and the processor implements the following steps when executing the computer program:
[0015] Determine a target image and at least two non-target images other than the target image from among multiple images to be screened; for each of the images to be screened, perform feature extraction on the images to be screened based on multiple feature extraction dimensions to obtain image features of the images to be screened corresponding to each of the feature extraction dimensions; based on each of the image features, perform feature similarity matching on the target image and each of the non-target images to obtain a similarity set between the target image and each of the non-target images; the similarity set includes feature similarities corresponding to each of the feature extraction dimensions; for each of the non-target images, perform statistics on multiple feature similarities corresponding to the non-target image to determine image similarity between the non-target image and the target image; retain the target image and non-target images whose image similarity is less than or equal to a similarity threshold.
[0016] In a fourth aspect, the present application further provides a computer-readable storage medium. The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the following steps are implemented:
[0017] Determine a target image and at least two non-target images other than the target image from among multiple images to be screened; for each of the images to be screened, perform feature extraction on the images to be screened based on multiple feature extraction dimensions to obtain image features of the images to be screened corresponding to each of the feature extraction dimensions; based on each of the image features, perform feature similarity matching on the target image and each of the non-target images to obtain a similarity set between the target image and each of the non-target images; the similarity set includes feature similarities corresponding to each of the feature extraction dimensions; for each of the non-target images, perform statistics on multiple feature similarities corresponding to the non-target image to determine image similarity between the non-target image and the target image; retain the target image and non-target images whose image similarity is less than or equal to a similarity threshold.
[0018] In a fifth aspect, the present application further provides a computer program product. The computer program product includes a computer program, and when the computer program is executed by a processor, the following steps are implemented:
[0019] Determine a target image and at least two non-target images other than the target image from among multiple images to be screened; for each of the images to be screened, perform feature extraction on the images to be screened based on multiple feature extraction dimensions to obtain image features of the images to be screened corresponding to each of the feature extraction dimensions; based on each of the image features, perform feature similarity matching on the target image and each of the non-target images to obtain a similarity set between the target image and each of the non-target images; the similarity set includes feature similarities corresponding to each of the feature extraction dimensions; for each of the non-target images, perform statistics on multiple feature similarities corresponding to the non-target image to determine image similarity between the non-target image and the target image; retain the target image and non-target images whose image similarity is less than or equal to a similarity threshold.
[0020] The above-mentioned image screening method, device, computer equipment, storage medium and computer program product determine a target image and at least two non-target images other than the target image from a plurality of images to be screened. For each image to be screened, based on a plurality of feature extraction dimensions, feature extraction is performed on the image to be screened respectively, and image features of the image to be screened corresponding to each feature extraction dimension are obtained. The extracted features are richer and are not limited to the extraction of features of a single dimension, which is conducive to improving the accuracy of subsequent image screening. Based on each image feature, feature similarity matching is performed on the target image and each non-target image respectively, and a similarity set of the target image and each non-target image is obtained. The similarity set includes feature similarities corresponding to each feature extraction dimension. For each non-target image, multiple feature similarities corresponding to the non-target image are statistically analyzed to determine the image similarity between the non-target image and the target image. By using the method, the obtained image similarity reflects the overall difference in image features between different images. The image similarity is more accurate than that of a single dimension, and the target image and non-target images whose image similarity is less than or equal to the similarity threshold are retained, thereby improving the accuracy of image screening. BRIEF DESCRIPTION OF THE DRAWINGS
[0021] FIG1 is a diagram showing an application environment of an image screening method according to an embodiment;
[0022] FIG2 is a schematic diagram of a flow chart of an image screening method in one embodiment;
[0023] FIG3 is a schematic diagram of a process of an industrial image data refining method based on self-supervised learning in one embodiment;
[0024] FIG4 is a schematic diagram of the structure of a speed codec in one embodiment;
[0025] FIG5 is a schematic diagram of the structure of a velocity coding block model in one embodiment;
[0026] FIG6 is a schematic diagram of the structure of a speed encoding module in one embodiment;
[0027] FIG7 is a schematic diagram of a structure diagram of a speed decoding module in one embodiment;
[0028] FIG8 is a schematic diagram of a first type of input and output images of a model in one embodiment;
[0029] FIG9 is a schematic diagram of a second type of input and output images of a model in one embodiment;
[0030] FIG10 is a schematic diagram of a third type of input and output images of a model in one embodiment;
[0031] FIG11 is a block diagram of an image screening device in one embodiment;
[0032] FIG. 12 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION
[0033] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.
[0034] The image screening method provided in the embodiment of the present application can be applied in the application environment shown in FIG1 . In which, the terminal 102 communicates with the server 104 via a network. The data storage system can store data that the server 104 needs to process. The data storage system can be integrated on the server 104, or it can be placed on the cloud or other network servers.
[0035] The terminal 102 determines a target image and at least two non-target images other than the target image from among the multiple images to be screened. For each image to be screened, the terminal 102 performs feature extraction on the image to be screened based on multiple feature extraction dimensions, and obtains image features of the image to be screened corresponding to each feature extraction dimension. Based on each image feature, the terminal 102 performs feature similarity matching on the target image and each non-target image, and obtains a similarity set between the target image and each non-target image, and the similarity set includes feature similarities corresponding to each feature extraction dimension. For each non-target image, the terminal 102 performs statistics on multiple feature similarities corresponding to the non-target image, and determines the image similarity between the non-target image and the target image. The terminal 102 retains the target image and non-target images whose image similarity is less than or equal to the similarity threshold.
[0036] The terminal 102 may obtain a plurality of images to be screened from the server 104. The terminal 102 may also send the retained target images and non-target images whose image similarity is less than or equal to a similarity threshold to the server 104.
[0037] The terminal 102 may be, but is not limited to, various personal computers, laptops, smart phones, tablet computers, IoT devices, and portable wearable devices. The IoT devices may be smart speakers, smart TVs, smart air conditioners, smart car-mounted devices, etc. The portable wearable devices may be smart watches, smart bracelets, head-mounted devices, etc. The server 104 may be implemented as an independent server or a server cluster consisting of multiple servers.
[0038] In one embodiment, as shown in FIG. 2 , an image screening method is provided, which is described by taking the method applied to the terminal 102 in FIG. 1 as an example, including:
[0039] S202: Determine a target image and at least two non-target images other than the target image from a plurality of images to be screened.
[0040] The image to be screened may represent an image taken of a workpiece. The image to be screened may represent an image of the entire workpiece, an image of a part of the workpiece, or an image of the environment in which the workpiece is located. For example, the image to be screened may be an image of scratches on the surface of the workpiece, an image of a bolt foolproofing measurement point, or an image of an abnormal point on the wood grain surface of the workpiece, etc. The number of images to be screened may be in the hundreds of thousands or millions.
[0041] The target image may be an image with better image quality among the multiple images to be screened, or may be an image randomly determined from the multiple images to be screened, or may be an image shot for a target object among the multiple images to be screened. The target object may be, for example, a workpiece defect.
[0042] The non-target image may be an image that is adjacent to the target image in each of the images to be screened. For example, if the target image is an image of a surface defect of a workpiece, the non-target image may be an image of a bolt or nut in the workpiece. The image acquisition objects of the non-target image and the target image may be the same or similar.
[0043] Specifically, the terminal may perform image quality scoring on each image to be screened, and obtain image quality scores of each image to be screened. The terminal may determine the image to be screened with the highest image quality score as the target image, and may determine the images to be screened other than the target image as non-target images. The terminal may also randomly determine a target image from a plurality of images to be screened, and for each target image, determine at least two non-target images whose image acquisition object matches the target object from other images other than the target image.
[0044] In one embodiment, the terminal can score the image quality of each image to be screened according to the image clarity and image size of each image to be screened, and obtain the image quality score of each image to be screened. For example, if the image clarity is higher and the image size is closer to the preset size, the image quality score of the image to be screened is higher.
[0045] S204, for each image to be screened, based on multiple feature extraction dimensions, feature extraction is performed on the image to be screened respectively to obtain image features of the image to be screened corresponding to each feature extraction dimension respectively.
[0046] Among them, feature extraction refers to a process of converting high-dimensional data into low-dimensional data through mapping or transformation methods.
[0047] The feature extraction dimension may represent the dimension of the low-dimensional data obtained after mapping the high-dimensional data. By performing feature extraction on the images to be screened according to different feature extraction dimensions, low-dimensional data in different dimensions may be obtained.
[0048] The image feature may represent the feature corresponding to the low-dimensional data obtained by extracting the feature of the image to be screened. The image feature may be the color feature, texture feature, shape feature, spatial relationship feature, etc. of the image to be screened.
[0049] Exemplarily, the terminal performs feature extraction on the images to be filtered based on multiple feature extraction dimensions. The terminal may perform grayscale feature extraction on the images to be filtered according to a grayscale detection method, the terminal may perform gradient feature extraction on the images to be filtered according to a gradient detection method, or the terminal may perform edge feature extraction on the images to be filtered according to an edge detection method.
[0050] Specifically, the terminal can determine the adjacent images of each image to be screened in each screening image, the terminal can splice the image to be screened and the adjacent images to obtain a spliced image, and the terminal can extract features from the spliced image based on multiple feature extraction dimensions to obtain image features of the image to be screened corresponding to each feature extraction dimension. Among them, the adjacent image can be an image that is adjacent to the image to be screened. Specifically, the adjacent image and the image to be screened can be two images to be screened with overlapping image areas.
[0051] Specifically, the terminal may also perform image segmentation processing on the image to be screened to obtain a plurality of segmented images. For each feature extraction dimension, the terminal may determine a feature extraction parameter that matches the feature extraction dimension, and the terminal may perform feature extraction on each segmented image according to the feature extraction parameter to obtain the segmented image features of each segmented image, and the terminal may fuse the segmented image features to obtain the image features of the image to be screened corresponding to each feature extraction dimension.
[0052] S206 , based on each image feature, perform feature similarity matching on the target image and each non-target image respectively, to obtain a similarity set between the target image and each non-target image.
[0053] The similarity set may represent a set corresponding to the result of feature similarity matching between the target image and the non-target image. For each non-target image, the similarity set may include the feature similarity corresponding to each feature extraction dimension between the non-target image and the target image.
[0054] Specifically, for each feature extraction dimension, the terminal can perform feature similarity matching on the target image and the non-target image based on the image features of each of the target image and the non-target image under the feature extraction dimension, and obtain the feature similarity between the target image and the non-target image under the feature extraction dimension. That is to say, feature similarity matching can represent similarity matching of the image features of the target image and the non-target image under the same feature extraction dimension. Taking the feature extraction dimension including the first dimension and the second dimension as an example, the image feature of the target image under the first dimension is A1 and the image feature under the second dimension is B1, and the image feature of the non-target image under the first dimension is A2 and the image feature under the second dimension is B2. At this time, feature similarity matching can represent similarity matching of A1 and A2, and similarity matching of B1 and B2. The similarity set of the non-target image may include the matching result of similarity matching of A1 and A2, and the matching result of similarity matching of B1 and B2.
[0055] S208 , for each non-target image, performing statistics on the similarities of multiple features corresponding to the non-target image to determine the image similarity between the non-target image and the target image.
[0056] Among them, image similarity can represent the degree of similarity between the non-target image and the target image. The higher the image similarity, the more likely it is that the non-target image and the target image are the same image. For example, in the field of industrial image recognition, the images to be screened can be images taken by a camera, and each image to be screened can be the same image or images with a high degree of similarity. In order to improve the efficiency of image post-processing, the same image or images with a high degree of similarity can be reduced, that is, the efficiency of image post-processing can be improved by reducing the number of images in the sample.
[0057] There may be a positive correlation between feature similarity and image similarity, that is, the higher the feature similarity corresponding to the non-target image, the higher the image similarity.
[0058] Specifically, the terminal may collect statistics on multiple feature similarities corresponding to the non-target image, and determine the maximum value among the feature similarities as the image similarity between the non-target image and the target image.
[0059] Specifically, the terminal can also obtain the screening scene, and the terminal can determine the feature similarity of the screening scene matching as the image similarity between the non-target image and the target image. Among them, the screening scene can be the scene corresponding to each image to be screened. For example, a scene with strong lighting, a scene with complex workpiece placement, etc. The feature similarity of the scene matching with strong lighting can be the feature similarity corresponding to the grayscale detection dimension. The feature similarity of the scene matching with complex workpiece placement can be the feature similarity corresponding to the spatial relationship detection dimension.
[0060] Specifically, the terminal may also determine in advance the feature weights configured for each feature extraction dimension. The terminal may perform statistics on multiple feature similarities corresponding to the non-target image according to each feature weight, and determine the image similarity between the non-target image and the target image.
[0061] S210, retaining the target image and non-target images whose image similarity is less than or equal to a similarity threshold.
[0062] The similarity threshold may be a threshold configured in advance for image similarity. Image similarity less than or equal to the similarity threshold may indicate that the non-target image is not very similar to the target image. Image similarity greater than the similarity threshold may indicate that the non-target image is highly similar to the target image.
[0063] Specifically, the terminal may retain the target image and non-target images whose image similarity is less than or equal to the similarity threshold. The terminal may also filter out non-target images whose image similarity is greater than the similarity threshold.
[0064] For example, if the target image is a bolt image and the non-target image is a nut image or a bolt image (the angle of the bolt in the non-target image is inconsistent with the angle of the bolt in the target image), the image similarity between the bolt image that is the target image and the nut image that is the non-target image may be less than or equal to the similarity threshold, and the nut image may be retained. The image similarity between the bolt image that is the target image and the bolt image that is the non-target image may be greater than the similarity threshold, and the nut image may be screened out.
[0065] In the above-mentioned image screening method, a target image and at least two non-target images other than the target image are determined from multiple images to be screened. For each image to be screened, based on multiple feature extraction dimensions, feature extraction is performed on the image to be screened respectively, and image features of the image to be screened corresponding to each feature extraction dimension are obtained. The extracted features are richer and are not limited to the extraction of features of a single dimension, which is conducive to improving the accuracy of subsequent image screening. Based on each image feature, feature similarity matching is performed on the target image and each non-target image respectively, and a similarity set of the target image and each non-target image is obtained. The similarity set includes feature similarities corresponding to each feature extraction dimension. For each non-target image, multiple feature similarities corresponding to the non-target image are statistically analyzed to determine the image similarity between the non-target image and the target image. By using the method, the obtained image similarity reflects the overall difference in image features between different images. The image similarity is more accurate than that of a single dimension, and the target image and non-target images whose image similarity is less than or equal to the similarity threshold are retained, thereby improving the accuracy of image screening.
[0066] In one embodiment, determining a target image and at least two non-target images other than the target image from a plurality of images to be screened includes: acquiring a plurality of images to be screened, determining an image quality score of each image to be screened, determining the target image from a plurality of images to be screened whose image quality scores satisfy a quality condition, determining a plurality of candidate images from a plurality of images to be screened whose image quality scores do not satisfy the quality condition, and screening at least two non-target images whose image acquisition objects match the target image from each of the candidate images.
[0067] The plurality of images to be screened may be read in advance or in real time, and are not limited here and can be flexibly selected according to actual conditions.
[0068] The image quality score may indicate the image quality of the image to be screened. For example, the image quality may indicate the clarity of the image to be screened, may indicate the image size of the image to be screened, and the like.
[0069] The quality condition may represent a filter condition configured for image quality scoring, for example, a clarity threshold, or a standard image size.
[0070] The target image may represent an image to be screened whose image quality score satisfies the quality condition. Specifically, the image quality score satisfying the quality condition may mean that the clarity of the image to be screened is greater than or equal to the clarity threshold, or that the image size of the image to be screened is the image standard size, or both.
[0071] Similarly, the candidate image may represent an image to be screened whose image quality score does not satisfy the quality condition.
[0072] The image acquisition object can represent the workpiece or the defect on the workpiece, for example, a bolt, a nut, or a crack on the workpiece surface.
[0073] A non-target image indicates that an image acquisition object in a candidate image matches an image acquisition object in a target image. For example, if the target object in the target image is a bolt, at least two non-target images whose image acquisition object matches the target image may be a candidate image containing the bolt, an image of a nut that matches the bolt, or an image of a defect on the surface of the bolt.
[0074] There are multiple ways to screen the images to be screened based on the image quality score. Specifically, the terminal can obtain multiple images to be screened in real time, the terminal can determine the image quality score of each image to be screened, the terminal can determine the target image from the multiple images to be screened whose image quality scores meet the quality conditions, the terminal can determine multiple candidate images from the multiple images to be screened whose image quality scores do not meet the quality conditions, and the terminal can screen at least two non-target images whose image acquisition objects match the target image from each candidate image.
[0075] Specifically, the terminal may also obtain multiple images to be screened, and the terminal may score the image quality of each image to be screened to obtain the image quality score of each image to be screened. The terminal may determine the image to be screened with the highest image quality score as the target image, and may determine the images to be screened other than the target image as non-target images. The terminal may also randomly determine the target image from the multiple images to be screened, and for each target image, determine at least two non-target images whose image acquisition object matches the target object from other images other than the target image.
[0076] In one embodiment, when the candidate images are obtained, the terminal may determine the positional relationship between the images to be screened based on the image information carried by each of the images to be screened. The terminal may screen multiple selected images adjacent to the target image position from the candidate images based on the positional relationship between the images to be screened. The terminal may screen at least two non-target images whose image acquisition objects match the target image from the selected images.
[0077] In this embodiment, multiple images to be screened are obtained, the image quality scores of each image to be screened are determined, a target image is determined from multiple images to be screened whose image quality scores meet the quality conditions, and multiple candidate images are determined from multiple images to be screened whose image quality scores do not meet the quality conditions. The target image has a higher quality score, which helps to improve the accuracy of subsequent image screening. At least two non-target images whose image acquisition objects match the target image are screened from each candidate image, and there is a strong correlation between the non-target image and the target image, which further improves the accuracy of the screened images.
[0078] In one embodiment, based on multiple feature extraction dimensions, feature extraction is performed on the images to be screened respectively to obtain image features of the images to be screened corresponding to each feature extraction dimension, including: performing image block processing on the images to be screened to obtain multiple block images, determining feature extraction parameters matching the feature extraction dimension for each feature extraction dimension, performing feature extraction on each block image according to the feature extraction parameters to obtain block image features of each block image, and fusing the features of each block image to obtain image features of the images to be screened corresponding to each feature extraction dimension.
[0079] The block image may be a partial area of the image to be screened. The block processing may represent uniform segmentation processing of the image to be screened.
[0080] The feature extraction parameters may represent model parameters of a model used to extract the segmented images. For example, an encoder may be used to extract features from the segmented images. The model parameters may be parameters in a trained encoder.
[0081] The block image feature represents the image feature of the block image, and the image feature can be expressed as a feature vector.
[0082] Fusion means fusing the features of the image blocks, for example, concatenating multiple feature vectors.
[0083] Specifically, the terminal can determine the position information of each block image in the image to be screened. Based on the position information of each block image and the features of the block images, the terminal obtains the image features of the image to be screened corresponding to each feature extraction dimension.
[0084] In one of the embodiments, for each feature extraction dimension, the terminal determines the weight of the block image in the fusion process based on the position information of the block image in the image to be filtered, and the terminal fuses the block image features of each block image according to the weight of each block image to obtain the image features of the image to be filtered corresponding to each feature extraction dimension.
[0085] In one of the embodiments, the fusion method can be to splice the feature vectors. For each feature extraction dimension, the terminal preferentially splices the block image features corresponding to the block images with higher weights, and then splices the block image features corresponding to the block images with lower weights, thereby obtaining the spliced image features, that is, the image features of each feature extraction dimension corresponding to the image to be screened.
[0086] In this embodiment, the image to be screened is divided into blocks to obtain a plurality of block images. For each feature extraction dimension, a feature extraction parameter matching the feature extraction dimension is determined. According to the feature extraction parameters, feature extraction is performed on each block image to obtain block image features of each block image. The features of each block image are fused to obtain image features of the image to be screened corresponding to each feature extraction dimension, thereby effectively reducing the size of the image to be screened and facilitating the improvement of the screening efficiency of large-size images.
[0087] In one embodiment, statistics are performed on multiple feature similarities corresponding to non-target images to determine image similarity between the non-target image and the target image, including: determining feature weights configured for each feature extraction dimension, and performing statistics on multiple feature similarities corresponding to the non-target image according to each feature weight to determine image similarity between the non-target image and the target image.
[0088] Among them, the feature extraction dimension can specifically be a grayscale detection dimension, a gradient detection dimension, a color detection dimension, an edge detection dimension, a texture detection dimension, a shape detection dimension, a spatial relationship detection dimension, etc. Different dimensions have different degrees of influence on the similarity between images. For example, in a scene with strong lighting, the grayscale detection dimension has a greater influence on the similarity between images. For another example, in a scene with weak lighting, the color detection dimension has a greater influence on the similarity between images. For another example, in a scene with complex placement of artifacts, the spatial relationship detection dimension has a greater influence on the similarity between images.
[0089] The feature weight can indicate the influence of the extracted image features corresponding to the feature extraction dimension on the similarity between images. The higher the feature weight, the higher the influence, and vice versa, the lower the feature weight, the lower the influence.
[0090] Specifically, the terminal can determine the feature weights configured for each feature extraction dimension according to the scene corresponding to each image to be screened. Exemplarily, the terminal can determine the feature weights configured for each feature extraction dimension according to factors such as the average light intensity in the scene generated by the image to be screened, the complexity of the workpiece placement, etc.
[0091] In this embodiment, the feature weights configured for each feature extraction dimension are determined, and according to each feature weight, the multiple feature similarities corresponding to the non-target image are statistically analyzed to determine the image similarity between the non-target image and the target image. According to different feature extraction dimension pairs, the accuracy of image similarity is improved.
[0092] In one embodiment, the image screening method also includes: determining a new target image and at least two new non-target images other than the new target image from among the retained multiple non-target images, determining, for each new non-target image, a new image similarity between the new non-target image and the new target image, retaining the new target image and the new non-target images whose new image similarity is less than or equal to a similarity threshold.
[0093] The new target image may be an image with better image quality among multiple non-target images.
[0094] The new non-target image may be an image in each non-target image that is adjacent to the new target image. For example, if the new target image is a surface defect image of a workpiece, the new non-target image may be an image of a bolt or a nut in the workpiece.
[0095] The new image similarity can indicate the degree of similarity between the new non-target image and the new target image. The higher the new image similarity, the more likely it is that the new non-target image and the new target image are the same image. To further improve the efficiency of image post-processing, the number of images that are the same or very similar can be reduced, that is, the efficiency of image post-processing can be improved by reducing the number of images in the sample.
[0096] Specifically, based on each image feature, the terminal performs feature similarity matching on the new target image and each new non-target image, and obtains a new similarity set between the new target image and each new non-target image, and the new similarity set includes a new feature similarity corresponding to each feature extraction dimension. For each new non-target image, the terminal performs statistics on multiple new feature similarities corresponding to the new non-target image, and determines the new image similarity between the new non-target image and the new target image.
[0097] In some other embodiments, the terminal retains the new image and the new non-target image whose similarity to the new image is less than or equal to the similarity threshold. The terminal screens out the new non-target image whose similarity to the new image is greater than the similarity threshold.
[0098] In this embodiment, a new target image and at least two new non-target images other than the new target image are determined from the multiple retained non-target images, and for each new non-target image, the new image similarity between the new non-target image and the new target image is determined, and the new target image and new non-target images whose new image similarity is less than or equal to a similarity threshold are retained, thereby further improving the efficiency of image post-processing by reducing the number of images in the sample.
[0099] In one embodiment, the image screening method also includes: constructing positive samples based on the target image and non-target images whose image similarity is greater than a similarity threshold, and constructing negative samples based on the non-target images whose image similarity is less than or equal to the similarity threshold, and the positive samples and negative samples are used to train an object recognition model; the object recognition model is used to identify the image acquisition object of the target image.
[0100] Among them, the non-target image whose image similarity is greater than the similarity threshold represents an image with greater image similarity with the target image.
[0101] Positive samples can represent samples of a certain target category, and negative samples represent samples that do not belong to a certain target category.
[0102] The object recognition model may refer to a model for recognizing an image acquisition object of a target image.
[0103] There are many ways to construct positive samples. Specifically, the terminal obtains a positive label configured for the target image. The terminal obtains a negative label configured for a non-target image whose image similarity is less than or equal to a similarity threshold. The terminal constructs a positive sample based on the positive label, the target image, and the non-target image whose image similarity is greater than the similarity threshold. The terminal constructs a negative sample based on the negative label and the non-target image whose image similarity is less than or equal to the similarity threshold.
[0104] Specifically, the terminal may perform binary classification on the target image and the non-target image whose image similarity is greater than the similarity threshold according to the target detection algorithm, and determine the image area classified as the foreground as a positive sample. The terminal may perform binary classification on the non-target image whose image similarity is less than or equal to the similarity threshold according to the target detection algorithm, and determine the image area classified as the background as a negative sample.
[0105] There are many ways to train the object recognition model. Specifically, the terminal trains the object recognition model with positive samples and negative samples to obtain a trained object recognition model.
[0106] In this embodiment, positive samples are constructed based on the target image and non-target images whose image similarity is greater than the similarity threshold. Negative samples are constructed based on non-target images whose image similarity is less than or equal to the similarity threshold. Positive samples and negative samples are used to train an object recognition model, which is used to identify the image acquisition object of the target image, and obtain an accurate object recognition model, which helps to improve the accuracy of subsequent image post-processing.
[0107] In one embodiment, as shown in FIG3 , a method for refining industrial image data based on self-supervised learning is provided, comprising:
[0108] S302, obtaining a plurality of images to be screened.
[0109] S304, determining the image quality score of each image to be screened.
[0110] S306, determining a target image from a plurality of images to be screened whose image quality scores meet the quality condition.
[0111] S308, determining a plurality of candidate images from a plurality of images to be screened whose image quality scores do not satisfy the quality condition.
[0112] S310, screening at least two non-target images whose image acquisition objects match the target image from each candidate image.
[0113] S312, for each image to be screened, performing image block processing on the image to be screened to obtain a plurality of block images.
[0114] S314: For each feature extraction dimension, determine a feature extraction parameter that matches the feature extraction dimension.
[0115] S316, performing feature extraction on each block image according to the feature extraction parameters to obtain the block image features of each block image.
[0116] Exemplarily, the feature extraction parameters may be parameters of the encoder in the speed encoding and decoding model. As shown in FIG4 , the structure diagram of the speed encoder-decoder, the speed encoder-decoder model (SED-VITM) function includes: inputting the original image into the encoder structure, and then inputting the output result of the encoder structure into the decoding structure, and the decoding structure outputs the Speed Encoder-Decoder (SED) image. Among them, the encoder structure is Speed Encoder Visioin Transformer, and the decoder structure is Speed Decoder Visioin Transformer. Among them, the feature extraction parameters may be parameters in the encoder structure.
[0117] The encoder structure includes an image block encoding module and an attention mechanism module. The image block encoding module may include a speed encoding block module. The structure of the speed encoding block model (Patch Encoder) is shown in Figure 5. The speed encoding block model divides the input image into N*N blocks of uniform size.
[0118] The image block encoding module can also include a speed encoding module. The structure of the speed encoding module (Speed VITEncoder) is shown in Figure 6. After the image is split into N*N patch images, it is sent to the speed encoding module for feature extraction to obtain a high-dimensional feature vector.
[0119] The Speed VIT Encoder is mainly composed of an Attention module, an MLP module and some BN layers. The Attention module is responsible for calculating the internal similarity of the encoded image and mapping the encoded patch image into a Query, Key, Value matrix.
[0120] The decoder structure takes the high-dimensional feature vector output by the encoder as input, outputs a simulated image of the original image size, and then calculates the difference between the original image and the simulated image to update the network. The decoder structure includes a speed decoding module, as shown in Figure 7, which is a structure diagram of the speed decoding module (Speed VIT Decoder). The speed decoding module decodes the high-dimensional features and restores them to the original size of the simulated image.
[0121] The speed decoding module is similar to the speed encoding module and is also composed of the Attention module, the MLP module, and the BN module. The speed decoding module randomly initializes the output and queries similar pixels based on the Query, Key, and Value to fill the output.
[0122] After obtaining the corresponding output image, the network output image and the original image can be used to calculate the loss, and the difference between the two images can be calculated to obtain the loss value to update the network. When the network training tends to be stable, the speed encoding module can be used as a feature extraction module.
[0123] S318, fusing the image features of each block to obtain the image features of the image to be screened corresponding to each feature extraction dimension.
[0124] S320 , based on each image feature, perform feature similarity matching on the target image and each non-target image respectively, to obtain a similarity set between the target image and each non-target image.
[0125] The similarity set includes the feature similarity corresponding to each feature extraction dimension.
[0126] S322: For each non-target image, determine the feature weights configured for each feature extraction dimension.
[0127] S324, performing statistics on the similarities of multiple features corresponding to the non-target image according to the feature weights, and determining the image similarity between the non-target image and the target image.
[0128] S326, retaining the target image and non-target images whose image similarity is less than or equal to the similarity threshold.
[0129] S328: Determine a new target image and at least two new non-target images other than the new target image from the plurality of retained non-target images.
[0130] S330: For each new non-target image, determine the new image similarity between the new non-target image and the new target image.
[0131] S332, retaining the new target image and the new non-target image whose similarity to the new image is less than or equal to the similarity threshold.
[0132] S334: construct positive samples based on the target image and non-target images whose image similarity is greater than a similarity threshold.
[0133] S336: construct negative samples based on non-target images whose image similarity is less than or equal to the similarity threshold.
[0134] The positive samples and negative samples are used to train an object recognition model. The object recognition model is used to recognize the image acquisition object of the target image.
[0135] For example, the object recognition model can be a Speed Encoder-Decoder (SED) model, which can accurately obtain the edge SED map required for measurement. In addition, expandable branches are added so that the model can not only output the SED map, but also obtain the measurement point category, angle and border.
[0136] Among them, as shown in Figure 8, the first column on the left represents three similar images in the massive data. When manually screening data, it often takes a long time to observe the details when encountering such difficult-to-distinguish images. It is very easy to classify these three images into one category. However, the efficiency of processing three images from the perspective of the trained model is 30 times that of manual work. The second column shows the binarization result of the feature map after the original image is input into the speed encoding and decoding model. It can be seen that the features are very easy to distinguish, some are linear, and some are clumpy. The rightmost column is the original output feature map output by the encoder in the speed encoding and decoding model. It can be seen that the activation area (with large feature values) is concentrated at the defect, and the details are very obvious. Features and images can be automatically distinguished by calculating the similarity and distance of various types of features. It can be concluded that the feature distinction based on the SED model for data distinction and refinement is much higher in accuracy and efficiency than manual screening.
[0137] Figure 9 shows a schematic diagram of the second type of input and output images of the model. The first column on the left represents three sample images of stud measurement points in the massive data. The task is to detect missing studs and scratches. The third image is very different from the first two, so manual screening is relatively easy. However, manual screening of the first and second images is very difficult, and the subjective factors are very strong. However, based on model feature extraction and similarity calculation, the situation of data differentiation is very obvious. The third column is the feature map extracted by the encoder in the model speed encoding and decoding model, and the second column is the feature vector obtained after the feature map is binary encoded. Based on the position and distance of different feature vectors, the data can be automatically classified and screened, which not only greatly improves the screening efficiency, but also eliminates the subjectivity of different manual screening, and achieves consistent data screening based on objective rules.
[0138] As shown in Figure 10, the third type of input and output images of the model are schematically shown. Figure 10 can be applied to the scene of wood grain surface crack detection. The first column on the left represents three wood grain surface sample images in the massive data. There is a problem of strong subjectivity in the manual screening process. It is too difficult to classify inconspicuous cracks. In addition, cracks are not obvious on the surface. Manual screening is extremely time-consuming and the effect is poor. The high-dimensional feature map in the third column can be the output image of the encoder in the speed coding and decoding model. It can be seen that the activation areas are all at the cracks. Binarizing the activation areas of the feature map can obtain a binary high-dimensional feature map. By calculating the distance and similarity between different features, the features can be quickly and accurately distinguished to achieve automatic data refinement.
[0139] In this embodiment, a target image and at least two non-target images other than the target image are determined from multiple images to be screened. For each image to be screened, based on multiple feature extraction dimensions, feature extraction is performed on the image to be screened respectively, and image features of the image to be screened corresponding to each feature extraction dimension are obtained. The extracted features are richer and are not limited to the extraction of features of a single dimension, which is conducive to improving the accuracy of subsequent image screening. Based on each image feature, feature similarity matching is performed on the target image and each non-target image respectively, and a similarity set of the target image and each non-target image is obtained. The similarity set includes feature similarities corresponding to each feature extraction dimension. For each non-target image, multiple feature similarities corresponding to the non-target image are statistically analyzed to determine the image similarity between the non-target image and the target image. By using the method, the obtained image similarity reflects the overall difference in image features between different images. The image similarity is more accurate than that of a single dimension, and the target image and non-target images whose image similarity is less than or equal to the similarity threshold are retained, thereby improving the accuracy of image screening. In this method, the sample is first input into the encoder for downsampling and encoding, and then upsampled and decoded by the decoder to obtain the output image and the original image to calculate the loss update network, and use the trained encoder network as the feature extractor. This can solve the problems of manual image screening, inability to distinguish different image features, and low efficiency. Secondly, for the Speed self-supervised model, an image segmentation module is added to accelerate training and better train the network receptive field.
[0140] It should be understood that, although the various steps in the flowcharts involved in the above-mentioned embodiments are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-mentioned embodiments can include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.
[0141] Based on the same inventive concept, the embodiment of the present application also provides an image screening device for implementing the above-mentioned image screening method. The implementation solution provided by the device to solve the problem is similar to the implementation solution recorded in the above-mentioned method, so the specific limitations in one or more image screening device embodiments provided below can refer to the limitations on the image screening method above, and will not be repeated here.
[0142] In one embodiment, as shown in FIG. 11 , an image screening device is provided, including: an image determination module 1102, a feature extraction module 1104, a feature matching module 1106, a statistics module 1108 and a retention module 1110, wherein:
[0143] An image determination module 1102 is used to determine a target image and at least two non-target images other than the target image from among the multiple images to be screened;
[0144] The feature extraction module 1104 is used to extract features of each image to be screened based on multiple feature extraction dimensions to obtain image features of the image to be screened corresponding to each feature extraction dimension;
[0145] The feature matching module 1106 is used to perform feature similarity matching on the target image and each non-target image based on each image feature, and obtain a similarity set between the target image and each non-target image; the similarity set includes the feature similarity corresponding to each feature extraction dimension;
[0146] A statistical module 1108 is used to perform statistics on the similarities of multiple features corresponding to each non-target image to determine the image similarity between the non-target image and the target image;
[0147] The retaining module 1110 is used to retain the target image and non-target images whose image similarity is less than or equal to a similarity threshold.
[0148] In one embodiment, the image determination module 1102 is further used to obtain multiple images to be screened; determine the image quality score of each image to be screened; determine a target image from the multiple images to be screened whose image quality scores meet the quality conditions; determine multiple candidate images from the multiple images to be screened whose image quality scores do not meet the quality conditions; and screen at least two non-target images whose image acquisition objects match the target image from each candidate image.
[0149] In one embodiment, the feature extraction module 1104 is also used to perform image segmentation processing on the image to be screened to obtain multiple segmented images; for each feature extraction dimension, determine the feature extraction parameters that match the feature extraction dimension; according to the feature extraction parameters, perform feature extraction on each segmented image to obtain the segmented image features of each segmented image; fuse the features of each segmented image to obtain the image features of the image to be screened corresponding to each feature extraction dimension.
[0150] In one embodiment, the statistical module 1108 is further used to determine the feature weights configured for each feature extraction dimension; according to each feature weight, statistics are performed on multiple feature similarities corresponding to the non-target image to determine the image similarity between the non-target image and the target image.
[0151] In one embodiment, the image screening device also includes: a new retention module, used to determine a new target image and at least two new non-target images other than the new target image among the multiple retained non-target images; for each new non-target image, determine the new image similarity between the new non-target image and the new target image; retain the new target image and the new non-target images whose new image similarity is less than or equal to a similarity threshold.
[0152] In one embodiment, the image screening device also includes: a construction module, which is used to construct positive samples based on the target image and non-target images whose image similarity is greater than a similarity threshold; construct negative samples based on non-target images whose image similarity is less than or equal to the similarity threshold; positive samples and negative samples are used to train an object recognition model; and the object recognition model is used to identify the image acquisition object of the target image.
[0153] Each module in the above-mentioned image screening device can be implemented in whole or in part by software, hardware or a combination thereof. Each module can be embedded in or independent of a processor in a computer device in the form of hardware, or can be stored in a memory in a computer device in the form of software, so that the processor can call and execute the operations corresponding to each module.
[0154] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be shown in FIG12. The computer device includes a processor, a memory, an input / output interface (Input / Output, referred to as I / O) and a communication interface. The processor, the memory and the input / output interface are connected via a system bus, and the communication interface is connected to the system bus via the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store image feature data. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, an image screening method is implemented.
[0155] Those skilled in the art will understand that the structure shown in FIG. 12 is merely a block diagram of a partial structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different arrangement of components.
[0156] In one embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and the above method steps are implemented when the processor executes the computer program.
[0157] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the above method steps are implemented.
[0158] In one embodiment, a computer program product is provided, comprising a computer program, which implements the above method steps when executed by a processor.
[0159] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited to this. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but are not limited to this.
[0160] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0161] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.
Claims
1. An image screening method, characterized in that: The method comprises: Determining a target image and at least two non-target images other than the target image from a plurality of images to be screened; For each of the images to be screened, based on a plurality of feature extraction dimensions, feature extraction is performed on the images to be screened respectively, to obtain image features of the images to be screened corresponding to each of the feature extraction dimensions respectively; Based on each of the image features, feature similarity matching is performed on the target image and each of the non-target images to obtain a similarity set between the target image and each of the non-target images; the similarity set includes feature similarities corresponding to each of the feature extraction dimensions; For each of the non-target images, statistics are collected on a plurality of feature similarities corresponding to the non-target image to determine an image similarity between the non-target image and the target image; The target image and non-target images whose image similarity is less than or equal to a similarity threshold are retained.
2. The method according to claim 1, characterized in that The step of determining a target image and at least two non-target images other than the target image from among the plurality of images to be screened comprises: Acquire multiple images to be screened; Determining the image quality score of each of the images to be screened; Determining a target image from a plurality of images to be screened whose image quality scores meet a quality condition; Determining a plurality of candidate images from a plurality of images to be screened whose image quality scores do not satisfy the quality condition; At least two non-target images whose image acquisition objects match the target image are screened from each of the candidate images.
3. The method according to claim 1, characterized in that The extracting features of the image to be screened based on multiple feature extraction dimensions to obtain image features of the image to be screened corresponding to each feature extraction dimension respectively includes: Performing image block processing on the image to be screened to obtain a plurality of block images; For each feature extraction dimension, determining a feature extraction parameter matching the feature extraction dimension; According to the feature extraction parameters, feature extraction is performed on each block image to obtain the block image features of each block image; The image features of each of the block images are fused to obtain image features of the image to be screened corresponding to each of the feature extraction dimensions.
4. The method according to claim 1, characterized in that: The performing statistics on the similarities of the multiple features corresponding to the non-target image to determine the image similarity between the non-target image and the target image includes: Determine the feature weights configured for each of the feature extraction dimensions; According to each of the feature weights, statistics are collected on a plurality of feature similarities corresponding to the non-target image to determine the image similarity between the non-target image and the target image.
5. The method according to claim 1, characterized in that The method further comprises: Determine a new target image and at least two new non-target images other than the new target image from the retained multiple non-target images; For each of the new non-target images, determining a new image similarity between the new non-target image and the new target image; The new target image and new non-target images whose similarity to the new image is less than or equal to the similarity threshold are retained.
6. The method according to any one of claims 1 to 5, characterized in that The method further comprises: constructing a positive sample based on the target image and a non-target image whose image similarity is greater than the similarity threshold; Constructing negative samples based on non-target images whose image similarity is less than or equal to the similarity threshold; The positive samples and the negative samples are used to train an object recognition model; the object recognition model is used to recognize the image acquisition object of the target image.
7. An image screening device, characterized in that: The device comprises: An image determination module, used to determine a target image and at least two non-target images other than the target image among a plurality of images to be screened; A feature extraction module is used to extract features of each of the images to be screened based on multiple feature extraction dimensions to obtain image features of the images to be screened corresponding to each of the feature extraction dimensions; A feature matching module, configured to perform feature similarity matching on the target image and each of the non-target images based on the image features, to obtain a similarity set between the target image and each of the non-target images; the similarity set includes feature similarities corresponding to each of the feature extraction dimensions; A statistical module, for performing statistics on a plurality of feature similarities corresponding to each of the non-target images, and determining an image similarity between the non-target image and the target image; The retention module is used to retain the target image and the non-target images whose image similarity is less than or equal to a similarity threshold.
8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 are implemented.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.
10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.