Image classification method and system based on VIT model

Abstract

This invention provides an image classification method and system based on the VIT model. Based on real-time shooting information from a monitoring terminal, the method preprocesses the dynamic image stream generated by the monitoring terminal to obtain an effective dynamic image stream of the target area. This effective dynamic image stream is then framed into a static image sequence, and the original dynamic image stream is used for image restoration. Based on the VIT model, all static images within the static image sequence are identified to obtain the individual image features of each static image. This divides all static images into several static image sets, ensuring that all static images within each set have high similarity and correlation on the global screen. Based on the image data requirements of the client's task, at least a portion of the static image sets are selected and sent to the client. This fully utilizes the VIT model to perform refined identification and classification of several images generated in spatial and temporal proximity, improving the scalability and usability of the image data.

Description

Technical Field

[0001] This invention relates to the field of image processing, and more particularly to an image classification method and system based on the VIT model. Background Technology

[0002] Traditional Convolutional Neural Network (CNN) models can classify images in complex, nonlinear function spaces. However, when processing large images, they require multiple downsampling steps, which can easily lead to information loss. Furthermore, traditional CNNs primarily rely on local receptivity to process images, failing to fully utilize global contextual information to aid classification, increasing the workload and reducing efficiency. VisionTransformer (VIT), an image classification model based on the Transformer architecture, applies self-attention mechanisms to the image domain, enabling feature extraction and classification. Existing VIT-based models achieve classification by capturing global contextual information and the correlations between different regions of an image. However, these methods only classify a single image and its overall context, failing to fully utilize the VIT model's ability to finely identify and classify spatially and temporally adjacent images. This prevents the segmentation of large numbers of images into sets with corresponding attributes, hindering the accurate provision of image data to clients and reducing the scalability and usability of image data. Summary of the Invention

[0003] The purpose of this invention is to provide an image classification method and system based on the VIT model. Based on real-time shooting information from the monitoring terminal, the method preprocesses the dynamic image stream generated by the monitoring terminal to obtain an effective dynamic image stream of the target area. This effective dynamic image stream is then framed into a static image sequence, and the original dynamic image stream is used for image restoration. Based on the VIT model, all static images within the static image sequence are identified to obtain the individual image features of each static image. This divides all static images into several static image sets, ensuring that all static images within each set have a high degree of similarity and correlation on the global screen. Based on the image data requirements of the client's task, at least a portion of the static image sets are selected and sent to the client. The VIT model is fully utilized to perform refined identification and classification of several images generated in spatial and temporal proximity, dividing a large number of images into image sets with corresponding attributes. This provides accurate image data to the client and improves the scalability and usability of the image data.

[0004] This invention is achieved through the following technical solution:

[0005] Image classification methods based on the VIT model include:

[0006] Based on the real-time shooting information from the monitoring terminal, the dynamic image stream generated by the monitoring terminal is preprocessed to obtain an effective dynamic image stream of the target area; the effective dynamic image stream is then processed into frames to obtain a static image sequence.

[0007] Based on the VIT model, all static images in the static image sequence are identified to obtain the image graphic features of each static image; based on the image graphic features, all static images are divided into several static image sets.

[0008] The client performs task execution monitoring to obtain the image data requirement information of the client's task execution; based on the image data requirement information, at least a portion of the static image set is selected and sent to the client.

[0009] Optionally, based on the real-time shooting information from the monitoring terminal, the dynamic image stream generated by the monitoring terminal is preprocessed to obtain an effective dynamic image stream of the target area; the effective dynamic image stream is then segmented into frames to obtain a static image sequence, including:

[0010] Information on external interference during the process of the monitoring terminal shooting the target area is obtained, and based on the information on external interference, the time information of the shooting obstacle event during the shooting process of the monitoring terminal is determined;

[0011] Based on the occurrence time information of the shooting obstacle event, the image segments matching the time axis in the dynamic image stream generated by the monitoring terminal are subjected to noise reduction filtering preprocessing to obtain the effective dynamic image stream of the target area;

[0012] The effective dynamic image stream is segmented into frames to obtain several static images; based on the image recognizability of all static images, all static images are filtered and arranged to obtain a static image sequence.

[0013] Optionally, based on the image recognizability of all still images, all still images are filtered and arranged to obtain a still image sequence, including:

[0014] Perform grayscale processing on the static images to obtain a grayscale image corresponding to each static image;

[0015] Based on the preset coordinates of the monitored target area, the core target area and non-core area are segmented and obtained.

[0016] The gradient response value S of each pixel block in the core target region is obtained by using a horizontal convolution kernel with the Sobel operator for each pixel block in the core target region. x (x, y); where (x, y) represents the two-dimensional coordinates corresponding to each pixel block;

[0017] The gradient response value S of each pixel block in the core target region is obtained by using a vertical convolution kernel with the Sobel operator for each pixel block in the core target region. y (x, y);

[0018] Using the gradient response value S of each pixel block in the horizontal direction x (x, y) and the gradient response value S in the vertical direction y (x, y) Obtain the gradient magnitude corresponding to each pixel block, and use the gradient magnitude corresponding to each pixel block to obtain the average gradient G of the target region. obj ;

[0019] Retrieve the grayscale value corresponding to each pixel block contained in the non-target region, and use the grayscale value corresponding to each pixel block contained in the non-target region to obtain the grayscale standard deviation σ of the non-core region. bg ;

[0020] Using the average gradient G of the target region obj Standard deviation of gray level in non-core areas σ bg Obtain the dynamic resolution adjustment factor D;

[0021] The dynamic resolution adjustment coefficient is obtained using the following formula:

[0022]

[0023] Where R represents the dynamic adjustment factor for sharpness; w g and w σ G represents the average gradient of the target region, respectively. obj Standard deviation of gray level in non-core areas σ bg The corresponding weight value;

[0024] The core target area is divided into m×m blocks to obtain multiple target area blocks, with m=3 being the preferred number.

[0025] The gradient magnitude of each pixel block in each target region block is compared with a preset magnitude threshold, and the percentage of pixels whose gradient magnitude is greater than the preset magnitude threshold is obtained.

[0026] The detail spatial concentration of the core target region is obtained by using the pixel proportion value of each target region block;

[0027] The concentration of detail space is obtained by the following formula:

[0028]

[0029] Where C represents the concentration of detail space; a represents the total number of target region blocks; Di D represents the i-th element; c This represents the absolute difference between the maximum and minimum pixel percentage values ​​in the target region block; D represents the average pixel percentage value corresponding to all target region blocks.

[0030] Sharpness determination parameters are generated by dynamically adjusting the detail spatial concentration and sharpness coefficients.

[0031] The sharpness determination parameter is obtained by the following formula:

[0032]

[0033] Where Q represents the sharpness determination parameter;

[0034] The sharpness determination parameter is compared with a preset parameter threshold. If the sharpness determination parameter is lower than the preset parameter threshold, the recognizability of the static image is determined to be poor.

[0035] Optionally, based on the VIT model, all static images within the static image sequence are identified to obtain the image graphic features of each static image; based on the image graphic features, all static images are divided into several static image sets, including:

[0036] Based on the VIT model, all static images in the static image sequence are identified to obtain the graphic contour features of each sub-region of each static image. Based on the position information of each sub-region of each sub-region in the global static image and the graphic contour features of each sub-region, the graphic contour and position mapping relationship of the static image is constructed.

[0037] Based on the characteristics of the image outline and position mapping relationship of each static image, the global image outline distribution similarity between any two static images is determined; based on the global image outline distribution similarity, all static images are divided into several static image sets.

[0038] Optionally, the client's task execution is monitored to obtain the image data requirement information for the client's task execution; based on the image data requirement information, at least a portion of the static image set is filtered and sent to the client, including:

[0039] The client performs front-end task execution monitoring to obtain all task processes executed by the client in the front end, thereby determining the image data requirement information corresponding to the client's execution of the task process; wherein, the image data requirement information includes the image data content attribute information expected to be obtained by the client corresponding to the execution of the task process;

[0040] The image data requirement information is compared with the image graphic outline distribution information of each of the static image sets, and at least a portion of the static image sets that match the image data requirement information are selected and sent to the client.

[0041] Image classification systems based on the VIT model include:

[0042] The dynamic image stream preprocessing module is used to preprocess the dynamic image stream generated by the monitoring terminal based on the real-time shooting information of the monitoring terminal to obtain an effective dynamic image stream of the target area.

[0043] A static image sequence generation module is used to perform frame-segmentation processing on the effective dynamic image stream to obtain a static image sequence;

[0044] The image graphic feature recognition module is used to recognize all static images in the static image sequence based on the VIT model, and obtain the image graphic features of each static image.

[0045] The static image set division module is used to divide all static images into several static image sets based on the graphic features of the image;

[0046] The static image set filtering and sending module is used to monitor the client's task execution, obtain the image data requirement information of the client's task execution, and based on the image data requirement information, filter at least a portion of the static image set and send it to the client.

[0047] Optionally, the dynamic image stream preprocessing module is used to preprocess the dynamic image stream generated by the monitoring terminal based on the real-time shooting information of the monitoring terminal, to obtain an effective dynamic image stream of the target area, including:

[0048] Information on external interference during the process of the monitoring terminal shooting the target area is obtained, and based on the information on external interference, the time information of the shooting obstacle event during the shooting process of the monitoring terminal is determined;

[0049] Based on the occurrence time information of the shooting obstacle event, the image segments matching the time axis in the dynamic image stream generated by the monitoring terminal are subjected to noise reduction filtering preprocessing to obtain the effective dynamic image stream of the target area;

[0050] The static image sequence generation module is used to perform frame segmentation processing on the effective dynamic image stream to obtain a static image sequence, including:

[0051] The effective dynamic image stream is segmented into frames to obtain several static images; based on the image recognizability of all static images, all static images are filtered and arranged to obtain a static image sequence.

[0052] Optionally, based on the image recognizability of all still images, all still images are filtered and arranged to obtain a still image sequence, including:

[0053] Perform grayscale processing on the static images to obtain a grayscale image corresponding to each static image;

[0054] Based on the preset coordinates of the monitored target area, the core target area and non-core area are segmented and obtained.

[0055] The gradient response value S of each pixel block in the core target region is obtained by using a horizontal convolution kernel with the Sobel operator for each pixel block in the core target region. x (x, y); where (x, y) represents the two-dimensional coordinates corresponding to each pixel block;

[0056] The gradient response value S of each pixel block in the core target region is obtained by using a vertical convolution kernel with the Sobel operator for each pixel block in the core target region. y (x, y);

[0057] Using the gradient response value S of each pixel block in the horizontal direction x (x, y) and the gradient response value S in the vertical direction y (x, y) Obtain the gradient magnitude corresponding to each pixel block, and use the gradient magnitude corresponding to each pixel block to obtain the average gradient G of the target region. obj ;

[0058] Retrieve the grayscale value corresponding to each pixel block contained in the non-target region, and use the grayscale value corresponding to each pixel block contained in the non-target region to obtain the grayscale standard deviation σ of the non-core region. bg ;

[0059] Using the average gradient G of the target region obj Standard deviation of gray level in non-core areas σ bg Obtain the dynamic resolution adjustment factor D;

[0060] The dynamic resolution adjustment coefficient is obtained using the following formula:

[0061]

[0062] Where R represents the dynamic adjustment factor for sharpness; w g and w σ G represents the average gradient of the target region, respectively. obj Standard deviation of gray level in non-core areas σ bg The corresponding weight value;

[0063] The core target area is divided into m×m blocks to obtain multiple target area blocks, with m=3 being the preferred number.

[0064] The gradient magnitude of each pixel block in each target region block is compared with a preset magnitude threshold, and the percentage of pixels whose gradient magnitude is greater than the preset magnitude threshold is obtained.

[0065] The detail spatial concentration of the core target region is obtained by using the pixel proportion value of each target region block;

[0066] The concentration of detail space is obtained by the following formula:

[0067]

[0068] Where C represents the concentration of detail space; a represents the total number of target region blocks; D i D represents the i-th element; c This represents the absolute difference between the maximum and minimum pixel percentage values ​​in the target region block; D represents the average pixel percentage value corresponding to all target region blocks.

[0069] Sharpness determination parameters are generated by dynamically adjusting the detail spatial concentration and sharpness coefficients.

[0070] The sharpness determination parameter is obtained by the following formula:

[0071]

[0072] Where Q represents the sharpness determination parameter;

[0073] The sharpness determination parameter is compared with a preset parameter threshold. If the sharpness determination parameter is lower than the preset parameter threshold, the recognizability of the static image is determined to be poor.

[0074] Optionally, the image graphic feature recognition module is used to recognize all static images in the static image sequence based on the VIT model, to obtain the image graphic features of each static image, including:

[0075] Based on the VIT model, all static images in the static image sequence are identified to obtain the graphic contour features of each sub-region of each static image. Based on the position information of each sub-region of each sub-region in the global static image and the graphic contour features of each sub-region, the graphic contour and position mapping relationship of the static image is constructed.

[0076] The static image set division module is used to divide all static images into several static image sets based on the image graphic features, including:

[0077] Based on the characteristics of the image outline and position mapping relationship of each static image, the global image outline distribution similarity between any two static images is determined; based on the global image outline distribution similarity, all static images are divided into several static image sets.

[0078] Optionally, the static image set filtering and sending module is used to monitor the client's task execution to obtain the image data requirement information of the client's task execution; based on the image data requirement information, it filters at least a portion of the static image set and sends it to the client, including:

[0079] The client performs front-end task execution monitoring to obtain all task processes executed by the client in the front end, thereby determining the image data requirement information corresponding to the client's execution of the task process; wherein, the image data requirement information includes the image data content attribute information expected to be obtained by the client corresponding to the execution of the task process;

[0080] The image data requirement information is compared with the image graphic outline distribution information of each of the static image sets, and at least a portion of the static image sets that match the image data requirement information are selected and sent to the client.

[0081] Compared with the prior art, the present invention has the following beneficial effects:

[0082] The image classification method and system based on the VIT model provided in this application preprocesses the dynamic image stream generated by the monitoring terminal based on the real-time shooting information of the monitoring terminal to obtain the effective dynamic image stream of the target area. The effective dynamic image stream is then framed into a static image sequence, and the original dynamic image stream is restored. Based on the VIT model, all static images in the static image sequence are identified to obtain the image graphic features of each static image. This divides all static images into several static image sets, ensuring that all static images in each static image set have a high similarity and correlation in the global picture. Based on the image data requirements of the client's task, at least a portion of the static image sets are selected and sent to the client. The VIT model is fully utilized to perform fine identification and classification of several images generated in close proximity in space and time, dividing a large number of images into image sets with corresponding attributes, providing accurate image data to the client, and improving the scalability and availability of image data. Attached Figure Description

[0083] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0084] Figure 1 This is a flowchart illustrating the image classification method based on the VIT model provided by the present invention.

[0085] Figure 2 This is a schematic diagram of the structure of the image classification system based on the VIT model provided by the present invention. Detailed Implementation

[0086] To make the above-mentioned objectives, features, and advantages of this application more apparent and understandable, the specific embodiments of this application will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, it should be noted that, for ease of description, only the parts relevant to this application are shown in the accompanying drawings, not the entire structure. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of this application.

[0087] The terms “comprising” and “having”, and any variations thereof, used in this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the steps or units listed, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to such process, method, product, or apparatus.

[0088] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0089] Please see Figure 1 As shown, an embodiment of this application provides an image classification method based on the VIT model. This VIT model-based image classification method includes:

[0090] Based on the real-time shooting information from the monitoring terminal, the dynamic video stream generated by the monitoring terminal is preprocessed to obtain the effective dynamic video stream of the target area; the effective dynamic video stream is then processed into frames to obtain a static image sequence.

[0091] Based on the VIT model, all static images in the static image sequence are identified to obtain the image graphic features of each static image; based on the image graphic features, all static images are divided into several static image sets.

[0092] Monitor the client's task execution to obtain the image data requirement information for the client's task execution; based on the image data requirement information, filter at least a portion of the static image set and send it to the client.

[0093] The beneficial effects of the above embodiments are as follows: This image classification method based on the VIT model preprocesses the dynamic image stream generated by the monitoring terminal based on the real-time shooting information of the monitoring terminal to obtain the effective dynamic image stream of the target area. The effective dynamic image stream is then framed into a static image sequence, and the original dynamic image stream is repaired. Based on the VIT model, all static images within the static image sequence are identified to obtain the image graphic features of each static image. This divides all static images into several static image sets, ensuring that all static images within each set have a high degree of similarity and correlation on the global screen. Based on the image data requirements of the client's task, at least a portion of the static image sets are selected and sent to the client. The VIT model is fully utilized to perform refined identification and classification of several images generated in close proximity in space and time, dividing a large number of images into image sets with corresponding attributes, providing accurate image data to the client, and improving the scalability and usability of image data.

[0094] In another embodiment, based on the real-time shooting information from the monitoring terminal, the dynamic image stream generated by the monitoring terminal is preprocessed to obtain an effective dynamic image stream of the target area; the effective dynamic image stream is then subjected to frame segmentation to obtain a static image sequence, including:

[0095] Acquire information on external interference during the monitoring terminal's shooting of the target area, and based on this information, determine the timing of shooting obstacle events during the shooting process.

[0096] Based on the time information of the shooting obstacle event, the image segments in the dynamic image stream generated by the monitoring terminal that match the time axis are subjected to noise reduction filtering preprocessing to obtain the effective dynamic image stream of the target area.

[0097] The effective dynamic video stream is segmented into frames to obtain several static images; based on the image recognizability of all static images, all static images are filtered and arranged to obtain a static image sequence.

[0098] The beneficial effects of the above embodiments are that, in video surveillance, cameras and other monitoring devices periodically scan and capture images of a target area to obtain a dynamic image stream of that area. During the capture process, the monitoring device is inevitably affected by external vibrations and / or external electromagnetic signals, resulting in unclear images of the target area, leading to image shake and / or focus instability. The dynamic image segments generated by the monitoring device corresponding to the time of the capture obstacle event usually contain a lot of noise. Directly processing the entire dynamic image stream using the VIT model can easily lead to problems such as image feature recognition errors. To improve the accuracy of the subsequent VIT model's recognition of the dynamic image stream, it is necessary to repair the corresponding dynamic image segments within the dynamic image stream. Therefore, based on the time axis interval corresponding to the time of the capture obstacle event, noise reduction filtering preprocessing is performed on the image segments in the dynamic image stream generated by the monitoring device that match this time axis interval. This achieves local image segment image repair of the dynamic image stream, obtaining an effective dynamic image stream of the target area, thereby improving the image quality of the dynamic image stream and providing reliable raw image data for the VIT model's image classification. Furthermore, the effective dynamic video stream is segmented into frames to obtain several static images. Since the noise reduction and filtering preprocessing of the dynamic video stream cannot completely repair every static image frame in the dynamic video stream, in order to remove static images with low image quality, the image discernibility (i.e., image resolution) of each static image is also identified. If the image discernibility is less than a preset discernibility threshold, the corresponding static image is removed. Thus, all the static images remaining after the removal are arranged in order of generation time from earliest to latest to obtain a static image sequence, which facilitates the subsequent VIT model to identify and classify all static images in the complete static image sequence.

[0099] In another embodiment, based on the image recognizability of all still images, all still images are filtered and arranged to obtain a still image sequence, including:

[0100] Perform grayscale processing on the static images to obtain a grayscale image corresponding to each static image;

[0101] Based on the preset coordinates of the monitored target area, the core target area and non-core area are segmented and obtained.

[0102] The gradient response value S of each pixel block in the core target region is obtained by using a horizontal convolution kernel with the Sobel operator for each pixel block in the core target region. x (x, y); where (x, y) represents the two-dimensional coordinates corresponding to each pixel block;

[0103] The gradient response value S of each pixel block in the core target region is obtained by using a vertical convolution kernel with the Sobel operator for each pixel block in the core target region. y (x, y);

[0104] Using the gradient response value S of each pixel block in the horizontal direction x (x, y) and the gradient response value S in the vertical direction y (x, y) Obtain the gradient magnitude corresponding to each pixel block, and use the gradient magnitude corresponding to each pixel block to obtain the average gradient G of the target region. obj ;

[0105] Retrieve the grayscale value corresponding to each pixel block contained in the non-target region, and use the grayscale value corresponding to each pixel block contained in the non-target region to obtain the grayscale standard deviation σ of the non-core region. bg ;

[0106] Using the average gradient G of the target region obj Standard deviation of gray level in non-core areas σ bg Obtain the dynamic resolution adjustment factor D;

[0107] The dynamic resolution adjustment coefficient is obtained using the following formula:

[0108]

[0109] Where R represents the dynamic adjustment factor for sharpness; w g and w σ G represents the average gradient of the target region, respectively. obj Standard deviation of gray level in non-core areas σ bg The corresponding weight value;

[0110] The core target area is divided into m×m blocks to obtain multiple target area blocks, with m=3 being the preferred number.

[0111] The gradient magnitude of each pixel block in each target region block is compared with a preset magnitude threshold, and the percentage of pixels whose gradient magnitude is greater than the preset magnitude threshold is obtained.

[0112] The detail spatial concentration of the core target region is obtained by using the pixel proportion value of each target region block;

[0113] The concentration of detail space is obtained by the following formula:

[0114]

[0115] Where C represents the concentration of detail space; a represents the total number of target region blocks; Di D represents the i-th element; c This represents the absolute difference between the maximum and minimum pixel percentage values ​​in the target region block; D represents the average pixel percentage value corresponding to all target region blocks.

[0116] Sharpness determination parameters are generated by dynamically adjusting the detail spatial concentration and sharpness coefficients.

[0117] The sharpness determination parameter is obtained by the following formula:

[0118]

[0119] Where Q represents the sharpness determination parameter;

[0120] The sharpness determination parameter is compared with a preset parameter threshold. If the sharpness determination parameter is lower than the preset parameter threshold, the recognizability of the static image is determined to be poor.

[0121] The beneficial effects of the above embodiments are as follows: First, by using grayscale processing and segmenting the core target area and non-core areas, the system focuses on monitoring key targets, effectively isolates background interference in non-core areas, and ensures that the sharpness assessment revolves around the core target, avoiding interference from background noise on the judgment results. Second, by using the Sobel operator to calculate the horizontal and vertical gradient response values ​​of pixel blocks in the core target area, the system accurately captures the edge and detail features of the core target, providing objective gradient data support for the quantification of target sharpness. Simultaneously, a dynamic sharpness adjustment coefficient is constructed based on the average gradient of the target area and the grayscale standard deviation of the non-core area, comprehensively considering the richness of detail in the core target and the degree of background interference, enabling the coefficient to dynamically adapt to different... By analyzing the target-background characteristics of the image, the scene adaptability of sharpness assessment is improved. Simultaneously, by dividing the core region into m×m blocks and calculating the pixel ratio of each block, a detail spatial concentration is constructed, quantifying the spatial distribution concentration of core target details. This overcomes the limitation of single-gradient assessment in reflecting detail distribution characteristics, further refining the dimensions of sharpness assessment. Furthermore, by integrating the sharpness dynamic adjustment coefficient and detail spatial concentration to generate sharpness judgment parameters, the recognizability of static images can be accurately determined. This efficiently filters out images with acceptable recognizability to form a sequence, avoiding the inclusion of low-quality images due to insufficient target details or background interference, ensuring the overall recognizability and application reliability of the static image sequence.

[0122] In another embodiment, all static images within a static image sequence are identified based on the VIT model to obtain the image graphic features of each static image; based on the image graphic features, all static images are divided into several static image sets, including:

[0123] Based on the VIT model, all static images in the static image sequence are identified, and the graphic contour features of each sub-region of each static image are obtained. Based on the position information of each sub-region of each static image in the global static image and the graphic contour features of each sub-region of each static image, the mapping relationship between the graphic contour and position of the static image is constructed.

[0124] Based on the characteristics of the image contour and position mapping relationship of each static image, the global image contour distribution similarity between any two static images is determined; based on the global image contour distribution similarity, all static images are divided into several static image sets.

[0125] The beneficial effects of the above embodiments are that the VIT model can perform region recognition of images, that is, obtain the graphic contour features of all sub-regions within the global image frame. The region recognition of images using the VIT model is a conventional technique in this field and will not be described in detail here. Considering that different static images within a static image sequence may only have differences in graphic contours in some sub-regions, while the graphic contours of the remaining sub-regions are basically the same, in order to perform global image similarity recognition for all static images within the static image sequence, the VIT model is first used to recognize all static images within the static image sequence, obtaining the graphic contour features of each sub-region under each static image. This allows for accurate determination of the graphic contour features of each sub-region within the global frame of each static image. Then, based on the position information of each sub-region within the global frame of the static image and the graphic contour features of each sub-region, a mapping relationship between the graphic contours and positions of the static image is constructed. This allows for a two-dimensional distribution representation of all graphic contour features within the global frame of the static image, facilitating subsequent identification of the similarity between different static images from the perspective of the two-dimensional spatial distribution of graphic contours. Based on the characteristics of the image contour and position mapping relationship of each static image, the similarity of the image contours of any two static images corresponding to the same sub-region of the image is first determined. Then, the average similarity of the image contours of the two static images corresponding to all sub-regions of the image is determined. If the average similarity of the image contours exceeds a preset similarity threshold, the two static images are divided into the same static image set. In this way, all static images are divided into several static image sets to ensure that there is a high degree of image contour similarity between all static images in each static image set.

[0126] In another embodiment, the client's task execution is monitored to obtain image data requirement information for the client's task execution; based on the image data requirement information, at least a portion of the static image set is filtered and sent to the client, including:

[0127] The client performs front-end task execution monitoring to obtain all task processes executed by the client in the front end, thereby determining the image data requirement information corresponding to the client's task execution process; wherein, the image data requirement information includes the image data content attribute information expected to be obtained by the client's task execution process;

[0128] The image data requirement information is compared with the graphic outline distribution information of each static image set, and at least a portion of the static image sets that match the image data requirement information are selected and sent to the client.

[0129] The beneficial effects of the above embodiments are that when a client such as a computing terminal executes a corresponding task associated with a monitoring terminal at the front end, it needs to rely on image data from the monitoring terminal. The type of image data required by the client varies depending on the task being executed. To ensure that the client executes the corresponding task correctly and accurately at the front end, it is necessary to determine the actual image data requirements of the client for different task processes. Specifically, the client monitors the execution of front-end tasks to obtain all task processes executed by the client at the front end. Then, each task process is identified to obtain the image data content attribute information that the client expects to obtain for each task process. Furthermore, the image data requirement information is compared with the image graphic contour distribution information of each static image set to obtain the information content matching degree between the image data requirement information and the image graphic contour distribution information of each static image. If the information content matching degree exceeds a preset matching degree threshold, the corresponding static image set is sent to the client, providing the client with accurate image data and improving the scalability and availability of image data.

[0130] Please see Figure 2 As shown, an embodiment of this application provides an image classification system based on the VIT model. This VIT model-based image classification system includes:

[0131] The dynamic image stream preprocessing module is used to preprocess the dynamic image stream generated by the monitoring terminal based on the real-time shooting information of the monitoring terminal, so as to obtain an effective dynamic image stream of the target area.

[0132] The static image sequence generation module is used to perform frame-by-frame processing on the effective dynamic image stream to obtain a static image sequence;

[0133] The image graphic feature recognition module is used to identify all static images in a static image sequence based on the VIT model, and obtain the image graphic features of each static image.

[0134] The static image set division module is used to divide all static images into several static image sets based on the graphic features of the image.

[0135] The static image set filtering and sending module is used to monitor the client's task execution, obtain the image data requirement information of the client's task execution, and based on the image data requirement information, filter at least a portion of the static image set and send it to the client.

[0136] The beneficial effects of the above embodiments are as follows: This image classification system based on the VIT model preprocesses the dynamic image stream generated by the monitoring terminal based on the real-time shooting information of the monitoring terminal to obtain an effective dynamic image stream of the target area. The effective dynamic image stream is then framed into a static image sequence, and the original dynamic image stream is repaired. Based on the VIT model, all static images within the static image sequence are identified to obtain the image graphic features of each static image. This divides all static images into several static image sets, ensuring that all static images within each set have a high degree of similarity and correlation on the global screen. Based on the image data requirements of the client's task, at least a portion of the static image sets are selected and sent to the client. The VIT model is fully utilized to perform refined identification and classification of several images generated in close proximity in space and time, dividing a large number of images into image sets with corresponding attributes, providing accurate image data to the client, and improving the scalability and usability of image data.

[0137] In another embodiment, the dynamic image stream preprocessing module is used to preprocess the dynamic image stream generated by the monitoring terminal based on the real-time shooting information of the monitoring terminal, to obtain an effective dynamic image stream of the target area, including:

[0138] Acquire information on external interference during the monitoring terminal's shooting of the target area, and based on this information, determine the timing of shooting obstacle events during the shooting process.

[0139] Based on the time information of the shooting obstacle event, the image segments in the dynamic image stream generated by the monitoring terminal that match the time axis are subjected to noise reduction filtering preprocessing to obtain the effective dynamic image stream of the target area.

[0140] The static image sequence generation module is used to perform frame segmentation processing on the effective dynamic video stream to obtain a static image sequence, including:

[0141] The effective dynamic video stream is segmented into frames to obtain several static images; based on the image recognizability of all static images, all static images are filtered and arranged to obtain a static image sequence.

[0142] The beneficial effects of the above embodiments are that, in video surveillance, cameras and other monitoring devices periodically scan and capture images of a target area to obtain a dynamic image stream of that area. During the capture process, the monitoring device is inevitably affected by external vibrations and / or external electromagnetic signals, resulting in unclear images of the target area, leading to image shake and / or focus instability. The dynamic image segments generated by the monitoring device corresponding to the time of the capture obstacle event usually contain a lot of noise. Directly processing the entire dynamic image stream using the VIT model can easily lead to problems such as image feature recognition errors. To improve the accuracy of the subsequent VIT model's recognition of the dynamic image stream, it is necessary to repair the corresponding dynamic image segments within the dynamic image stream. Therefore, based on the time axis interval corresponding to the time of the capture obstacle event, noise reduction filtering preprocessing is performed on the image segments in the dynamic image stream generated by the monitoring device that match this time axis interval. This achieves local image segment image repair of the dynamic image stream, obtaining an effective dynamic image stream of the target area, thereby improving the image quality of the dynamic image stream and providing reliable raw image data for the VIT model's image classification. Furthermore, the effective dynamic video stream is segmented into frames to obtain several static images. Since the noise reduction and filtering preprocessing of the dynamic video stream cannot completely repair every static image frame in the dynamic video stream, in order to remove static images with low image quality, the image discernibility (i.e., image resolution) of each static image is also identified. If the image discernibility is less than a preset discernibility threshold, the corresponding static image is removed. Thus, all the static images remaining after the removal are arranged in order of generation time from earliest to latest to obtain a static image sequence, which facilitates the subsequent VIT model to identify and classify all static images in the complete static image sequence.

[0143] In another embodiment, based on the image recognizability of all still images, all still images are filtered and arranged to obtain a still image sequence, including:

[0144] Perform grayscale processing on the static images to obtain a grayscale image corresponding to each static image;

[0145] Based on the preset coordinates of the monitored target area, the core target area and non-core area are segmented and obtained.

[0146] The gradient response value S of each pixel block in the core target region is obtained by using a horizontal convolution kernel with the Sobel operator for each pixel block in the core target region. x (x, y); where (x, y) represents the two-dimensional coordinates corresponding to each pixel block;

[0147] The gradient response value S of each pixel block in the core target region is obtained by using a vertical convolution kernel with the Sobel operator for each pixel block in the core target region. y (x, y);

[0148] Using the gradient response value S of each pixel block in the horizontal direction x (x, y) and the gradient response value S in the vertical direction y (x, y) Obtain the gradient magnitude corresponding to each pixel block, and use the gradient magnitude corresponding to each pixel block to obtain the average gradient G of the target region. obj ;

[0149] Retrieve the grayscale value corresponding to each pixel block contained in the non-target region, and use the grayscale value corresponding to each pixel block contained in the non-target region to obtain the grayscale standard deviation σ of the non-core region. bg ;

[0150] Using the average gradient G of the target region obj Standard deviation of gray level in non-core areas σ bg Obtain the dynamic resolution adjustment factor D;

[0151] The dynamic resolution adjustment coefficient is obtained using the following formula:

[0152]

[0153] Where R represents the dynamic adjustment factor for sharpness; w g and w σ G represents the average gradient of the target region, respectively. obj Standard deviation of gray level in non-core areas σ bg The corresponding weight value;

[0154] The core target area is divided into m×m blocks to obtain multiple target area blocks, with m=3 being the preferred number.

[0155] The gradient magnitude of each pixel block in each target region block is compared with a preset magnitude threshold, and the percentage of pixels whose gradient magnitude is greater than the preset magnitude threshold is obtained.

[0156] The detail spatial concentration of the core target region is obtained by using the pixel proportion value of each target region block;

[0157] The concentration of detail space is obtained by the following formula:

[0158]

[0159] Where C represents the concentration of detail space; a represents the total number of target region blocks; Di D represents the i-th element; c This represents the absolute difference between the maximum and minimum pixel percentage values ​​in the target region block; D represents the average pixel percentage value corresponding to all target region blocks.

[0160] Sharpness determination parameters are generated by dynamically adjusting the detail spatial concentration and sharpness coefficients.

[0161] The sharpness determination parameter is obtained by the following formula:

[0162]

[0163] Where Q represents the sharpness determination parameter;

[0164] The sharpness determination parameter is compared with a preset parameter threshold. If the sharpness determination parameter is lower than the preset parameter threshold, the recognizability of the static image is determined to be poor.

[0165] The beneficial effects of the above embodiments are as follows: First, by using grayscale processing and segmenting the core target area and non-core areas, the system focuses on monitoring key targets, effectively isolates background interference in non-core areas, and ensures that the sharpness assessment revolves around the core target, avoiding interference from background noise on the judgment results. Second, by using the Sobel operator to calculate the horizontal and vertical gradient response values ​​of pixel blocks in the core target area, the system accurately captures the edge and detail features of the core target, providing objective gradient data support for the quantification of target sharpness. Simultaneously, a dynamic sharpness adjustment coefficient is constructed based on the average gradient of the target area and the grayscale standard deviation of the non-core area, comprehensively considering the richness of detail in the core target and the degree of background interference, enabling the coefficient to dynamically adapt to different... By analyzing the target-background characteristics of the image, the scene adaptability of sharpness assessment is improved. Simultaneously, by dividing the core region into m×m blocks and calculating the pixel ratio of each block, a detail spatial concentration is constructed, quantifying the spatial distribution concentration of core target details. This overcomes the limitation of single-gradient assessment in reflecting detail distribution characteristics, further refining the dimensions of sharpness assessment. Furthermore, by integrating the sharpness dynamic adjustment coefficient and detail spatial concentration to generate sharpness judgment parameters, the recognizability of static images can be accurately determined. This efficiently filters out images with acceptable recognizability to form a sequence, avoiding the inclusion of low-quality images due to insufficient target details or background interference, ensuring the overall recognizability and application reliability of the static image sequence.

[0166] In another embodiment, the image graphic feature recognition module is used to recognize all static images within a static image sequence based on the VIT model, obtaining the image graphic features of each static image, including:

[0167] Based on the VIT model, all static images in the static image sequence are identified, and the graphic contour features of each sub-region of each static image are obtained. Based on the position information of each sub-region of each static image in the global static image and the graphic contour features of each sub-region of each static image, the mapping relationship between the graphic contour and position of the static image is constructed.

[0168] The static image set partitioning module is used to divide all static images into several static image sets based on the graphic features of the image, including:

[0169] Based on the characteristics of the image contour and position mapping relationship of each static image, the global image contour distribution similarity between any two static images is determined; based on the global image contour distribution similarity, all static images are divided into several static image sets.

[0170] The beneficial effects of the above embodiments are that the VIT model can perform region recognition of images, that is, obtain the graphic contour features of all sub-regions within the global image frame. The region recognition of images using the VIT model is a conventional technique in this field and will not be described in detail here. Considering that different static images within a static image sequence may only have differences in graphic contours in some sub-regions, while the graphic contours of the remaining sub-regions are basically the same, in order to perform global image similarity recognition for all static images within the static image sequence, the VIT model is first used to recognize all static images within the static image sequence, obtaining the graphic contour features of each sub-region under each static image. This allows for accurate determination of the graphic contour features of each sub-region within the global frame of each static image. Then, based on the position information of each sub-region within the global frame of the static image and the graphic contour features of each sub-region, a mapping relationship between the graphic contours and positions of the static image is constructed. This allows for a two-dimensional distribution representation of all graphic contour features within the global frame of the static image, facilitating subsequent identification of the similarity between different static images from the perspective of the two-dimensional spatial distribution of graphic contours. Based on the characteristics of the image contour and position mapping relationship of each static image, the similarity of the image contours of any two static images corresponding to the same sub-region of the image is first determined. Then, the average similarity of the image contours of the two static images corresponding to all sub-regions of the image is determined. If the average similarity of the image contours exceeds a preset similarity threshold, the two static images are divided into the same static image set. In this way, all static images are divided into several static image sets to ensure that there is a high degree of image contour similarity between all static images in each static image set.

[0171] In another embodiment, the static image set filtering and sending module is used to monitor the client's task execution to obtain the image data requirement information for the client's task execution; based on the image data requirement information, it filters at least a portion of the static image set and sends it to the client, including:

[0172] The client performs front-end task execution monitoring to obtain all task processes executed by the client in the front end, thereby determining the image data requirement information corresponding to the client's task execution process; wherein, the image data requirement information includes the image data content attribute information expected to be obtained by the client's task execution process;

[0173] The image data requirement information is compared with the graphic outline distribution information of each static image set, and at least a portion of the static image sets that match the image data requirement information are selected and sent to the client.

[0174] The beneficial effects of the above embodiments are that when a client such as a computing terminal executes a corresponding task associated with a monitoring terminal at the front end, it needs to rely on image data from the monitoring terminal. The type of image data required by the client varies depending on the task being executed. To ensure that the client executes the corresponding task correctly and accurately at the front end, it is necessary to determine the actual image data requirements of the client for different task processes. Specifically, the client monitors the execution of front-end tasks to obtain all task processes executed by the client at the front end. Then, each task process is identified to obtain the image data content attribute information that the client expects to obtain for each task process. Furthermore, the image data requirement information is compared with the image graphic contour distribution information of each static image set to obtain the information content matching degree between the image data requirement information and the image graphic contour distribution information of each static image. If the information content matching degree exceeds a preset matching degree threshold, the corresponding static image set is sent to the client, providing the client with accurate image data and improving the scalability and availability of image data.

[0175] In summary, this image classification method and system based on the VIT model preprocesses the dynamic image stream generated by the monitoring terminal based on the real-time shooting information to obtain the effective dynamic image stream of the target area. This effective dynamic image stream is then framed into a static image sequence, and the original dynamic image stream is used for image restoration. Based on the VIT model, all static images within the static image sequence are identified to obtain the individual image features of each static image. This divides all static images into several static image sets, ensuring that all static images within each set have a high degree of similarity and correlation on the global screen. Based on the image data requirements of the client's task, at least a portion of the static image sets are selected and sent to the client. The VIT model is fully utilized to perform refined identification and classification of several images generated in spatial and temporal proximity, dividing a large number of images into image sets with corresponding attributes. This provides accurate image data to the client and improves the scalability and usability of the image data.

[0176] The above is only one specific embodiment of the present invention, and any improvements made based on the concept of the present invention shall be considered within the scope of protection of the present invention.

Claims

1. An image classification method based on the VIT model, characterized in that, include: Based on the real-time shooting information from the monitoring terminal, the dynamic image stream generated by the monitoring terminal is preprocessed to obtain an effective dynamic image stream of the target area; the effective dynamic image stream is then processed into frames to obtain a static image sequence. Based on the VIT model, all static images in the static image sequence are identified to obtain the image graphic features of each static image; based on the image graphic features, all static images are divided into several static image sets. The client performs task execution monitoring to obtain the image data requirement information for the client's task execution; based on the image data requirement information, at least a portion of the static image set is selected and sent to the client.

2. The image classification method based on the VIT model as described in claim 1, characterized in that: Based on the real-time shooting information from the monitoring terminal, the dynamic video stream generated by the monitoring terminal is preprocessed to obtain an effective dynamic video stream of the target area; the effective dynamic video stream is then subjected to frame segmentation to obtain a static image sequence, including: Information on external interference during the process of the monitoring terminal shooting the target area is obtained, and based on the information on external interference, the time information of the shooting obstacle event during the shooting process of the monitoring terminal is determined; Based on the time information of the shooting obstacle event, the image segments matching the time axis in the dynamic image stream generated by the monitoring terminal are subjected to noise reduction filtering preprocessing to obtain the effective dynamic image stream of the target area; The effective dynamic image stream is processed by frame segmentation to obtain several static images; based on the image recognizability of all static images, all static images are filtered and arranged to obtain a static image sequence.

3. The image classification method based on the VIT model as described in claim 2, characterized in that: Based on the image recognizability of all still images, all still images are filtered and arranged to obtain a still image sequence, including: Perform grayscale processing on the static images to obtain a grayscale image corresponding to each static image; Based on the preset coordinates of the monitored target area, the core target area and non-core area are segmented and obtained. The gradient response value S of each pixel block in the core target region is obtained by using a horizontal convolution kernel with the Sobel operator for each pixel block in the core target region. x (x, y); where (x, y) represents the two-dimensional coordinates corresponding to each pixel block; The gradient response value S of each pixel block in the core target region is obtained by using a vertical convolution kernel with the Sobel operator for each pixel block in the core target region. y (x, y); Using the gradient response value S of each pixel block in the horizontal direction x (x, y) and the gradient response value S in the vertical direction y (x, y) Obtain the gradient magnitude corresponding to each pixel block, and use the gradient magnitude corresponding to each pixel block to obtain the average gradient G of the target region. obj ; Retrieve the grayscale value of each pixel block contained in the non-target region, and use the grayscale value of each pixel block contained in the non-target region to obtain the grayscale standard deviation σ of the non-core region. bg ; Using the average gradient G of the target region obj Standard deviation of gray level in non-core areas σ bg Obtain the dynamic resolution adjustment factor D; The dynamic resolution adjustment coefficient is obtained using the following formula: Where R represents the dynamic adjustment factor for sharpness; w g and w σ G represents the average gradient of the target region, respectively. obj Standard deviation of gray level in non-core areas σ bg The corresponding weight value; The core target area is divided into m×m blocks to obtain multiple target area blocks; The gradient magnitude of each pixel block in each target region block is compared with a preset magnitude threshold, and the percentage of pixels whose gradient magnitude is greater than the preset magnitude threshold is obtained. The detail spatial concentration of the core target region is obtained by using the pixel proportion value of each target region block; The concentration of detail space is obtained by the following formula: Where C represents the concentration of detail space; a represents the total number of target region blocks; D i D represents the i-th element; c This represents the absolute difference between the maximum and minimum pixel percentage values ​​in the target region block; D represents the average pixel percentage value corresponding to all target region blocks. Sharpness determination parameters are generated by dynamically adjusting the detail spatial concentration and sharpness coefficients. The sharpness determination parameter is obtained by the following formula: Where Q represents the sharpness determination parameter; The sharpness determination parameter is compared with a preset parameter threshold. If the sharpness determination parameter is lower than the preset parameter threshold, the recognizability of the static image is determined to be poor.

4. The image classification method based on the VIT model as described in claim 1, characterized in that: Based on the VIT model, all static images in the static image sequence are identified to obtain the image graphic features of each static image. Based on the aforementioned graphic features, all static images are divided into several static image sets, including: Based on the VIT model, all static images in the static image sequence are identified to obtain the graphic contour features of each sub-region of each static image. Based on the position information of each sub-region of each sub-region in the global static image and the graphic contour features of each sub-region, the graphic contour and position mapping relationship of the static image is constructed. Based on the characteristics of the image outline and position mapping relationship of each static image, the global image outline distribution similarity between any two static images is determined; based on the global image outline distribution similarity, all static images are divided into several static image sets.

5. The image classification method based on the VIT model as described in claim 1, characterized in that: Monitor the client's task execution to obtain the image data requirement information for the client's task execution. Based on the image data requirement information, at least a portion of the static image set is selected and sent to the client, including: The client performs front-end task execution monitoring to obtain all task processes executed by the client in the front end, thereby determining the image data requirement information corresponding to the client's execution of the task process; wherein, the image data requirement information includes the image data content attribute information expected to be obtained by the client corresponding to the execution of the task process; The image data requirement information is compared with the image graphic outline distribution information of each of the static image sets, and at least a portion of the static image sets that match the image data requirement information are selected and sent to the client.

6. An image classification system based on the VIT model, characterized in that, include: The dynamic image stream preprocessing module is used to preprocess the dynamic image stream generated by the monitoring terminal based on the real-time shooting information of the monitoring terminal to obtain an effective dynamic image stream of the target area. A static image sequence generation module is used to perform frame-segmentation processing on the effective dynamic image stream to obtain a static image sequence; The image graphic feature recognition module is used to recognize all static images in the static image sequence based on the VIT model, and obtain the image graphic features of each static image. The static image set division module is used to divide all static images into several static image sets based on the graphic features of the image; The static image set filtering and sending module is used to monitor the client's task execution, obtain the image data requirement information of the client's task execution, and based on the image data requirement information, filter at least a portion of the static image set and send it to the client.

7. The image classification system based on the VIT model as described in claim 6, characterized in that: The dynamic image stream preprocessing module is used to preprocess the dynamic image stream generated by the monitoring terminal based on the real-time shooting information of the monitoring terminal, to obtain an effective dynamic image stream of the target area, including: Information on external interference during the process of the monitoring terminal shooting the target area is obtained, and based on the information on external interference, the time information of the shooting obstacle event during the shooting process of the monitoring terminal is determined; Based on the time information of the shooting obstacle event, the image segments matching the time axis in the dynamic image stream generated by the monitoring terminal are subjected to noise reduction filtering preprocessing to obtain the effective dynamic image stream of the target area; The static image sequence generation module is used to perform frame segmentation processing on the effective dynamic image stream to obtain a static image sequence, including: The effective dynamic image stream is processed by frame segmentation to obtain several static images; based on the image recognizability of all static images, all static images are filtered and arranged to obtain a static image sequence.

8. The image classification system based on the VIT model as described in claim 7, characterized in that: Based on the image recognizability of all still images, all still images are filtered and arranged to obtain a still image sequence, including: Perform grayscale processing on the static images to obtain a grayscale image corresponding to each static image; Based on the preset coordinates of the monitored target area, the core target area and non-core area are segmented and obtained. The gradient response value S of each pixel block in the core target region is obtained by using a horizontal convolution kernel with the Sobel operator for each pixel block in the core target region. x (x, y); where (x, y) represents the two-dimensional coordinates corresponding to each pixel block; The gradient response value S of each pixel block in the core target region is obtained by using a vertical convolution kernel with the Sobel operator for each pixel block in the core target region. y (x, y); Using the gradient response value S of each pixel block in the horizontal direction x (x, y) and the gradient response value S in the vertical direction y (x, y) Obtain the gradient magnitude corresponding to each pixel block, and use the gradient magnitude corresponding to each pixel block to obtain the average gradient G of the target region. obj ; Retrieve the grayscale value of each pixel block contained in the non-target region, and use the grayscale value of each pixel block contained in the non-target region to obtain the grayscale standard deviation σ of the non-core region. bg ; Using the average gradient G of the target region obj Standard deviation of gray level in non-core areas σ bg Obtain the dynamic resolution adjustment factor D; The dynamic resolution adjustment coefficient is obtained using the following formula: Where R represents the dynamic adjustment factor for sharpness; w g and w σ G represents the average gradient of the target region, respectively. obj Standard deviation of gray level in non-core areas σ bg The corresponding weight value; The core target area is divided into m×m blocks to obtain multiple target area blocks; The gradient magnitude of each pixel block in each target region block is compared with a preset magnitude threshold, and the percentage of pixels whose gradient magnitude is greater than the preset magnitude threshold is obtained. The detail spatial concentration of the core target region is obtained by using the pixel proportion value of each target region block; The concentration of detail space is obtained by the following formula: Where C represents the concentration of detail space; a represents the total number of target region blocks; D i D represents the i-th element; c This represents the absolute difference between the maximum and minimum pixel percentage values ​​in the target region block; D represents the average pixel percentage value corresponding to all target region blocks. Sharpness determination parameters are generated by dynamically adjusting the detail spatial concentration and sharpness coefficients. The sharpness determination parameter is obtained by the following formula: Where Q represents the sharpness determination parameter; The sharpness determination parameter is compared with a preset parameter threshold. If the sharpness determination parameter is lower than the preset parameter threshold, the recognizability of the static image is determined to be poor.

9. The image classification system based on the VIT model as described in claim 6, characterized in that: The image graphic feature recognition module is used to recognize all static images in the static image sequence based on the VIT model, and obtain the image graphic features of each static image, including: Based on the VIT model, all static images in the static image sequence are identified to obtain the graphic contour features of each sub-region of each static image. Based on the position information of each sub-region of each sub-region in the global static image and the graphic contour features of each sub-region, the graphic contour and position mapping relationship of the static image is constructed. The static image set division module is used to divide all static images into several static image sets based on the image graphic features, including: Based on the characteristics of the image outline and position mapping relationship of each static image, the global image outline distribution similarity between any two static images is determined; based on the global image outline distribution similarity, all static images are divided into several static image sets.

10. The image classification system based on the VIT model as described in claim 6, characterized in that: The static image set filtering and sending module is used to monitor the client's task execution and obtain the image data requirement information of the client's task execution. Based on the image data requirement information, at least a portion of the static image set is selected and sent to the client, including: The client performs front-end task execution monitoring to obtain all task processes executed by the client in the front end, thereby determining the image data requirement information corresponding to the client's execution of the task process; wherein, the image data requirement information includes the image data content attribute information expected to be obtained by the client corresponding to the execution of the task process; The image data requirement information is compared with the image graphic outline distribution information of each of the static image sets, and at least a portion of the static image sets that match the image data requirement information are selected and sent to the client.

Images