A method and system for detecting defects in a label paper

Abstract

The present application relates to label paper detection technical field, especially in kind of label paper's defect detection method and system, the method includes: obtaining a plurality of historical monitoring videos corresponding to the detection process and as a video to be identified, a plurality of preset to-be-identified objects in the video to be identified are behavior identified, and the spatial difference degree, the time difference degree and the sequence difference degree are obtained with the preset standard action video, the video abnormal score value corresponding to the video to be identified is calculated by combining the corresponding preset index weight, so as to screen out the behavior identification abnormal video, and then determine the corresponding a plurality of label paper to be detected, the target image of the label paper to be detected is recognized by using a deep learning model, the target defect type is obtained, and the label paper to be detected with defects is rejected; the present application can improve the probability of identifying the defective label paper, and also improve the detection efficiency and detection quality of the label paper defect.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of label paper detection, in particular to a label paper defect detection method and system. BACKGROUND

[0002] In a modern label paper production line, manual and automatic equipment usually work together, and manual intervention is mainly concentrated in specific links, such as material loading, mold installation, inspection and rejection of defective products, and adhesive viscosity testing. Problems in each link can cause defects in label paper. Due to the precision work of automatic equipment, the probability of label paper defects is low, while manual work may cause a large number of defective products due to work errors. In order to ensure the quality of label paper, the product needs to be rechecked, and since label paper is mostly packaged in stacks or rolls, manual inspection of all label paper will consume a lot of time, so sampling inspection is needed, but the sampled products are random, the probability of sampling defective label paper is low, and there is also a phenomenon of manual misidentification, which affects the sampling inspection effect. SUMMARY

[0003] In view of the above technical problems, the present application provides a label paper defect detection method and system, which can improve the probability of identifying defective label paper, and also improve the detection efficiency and detection quality of label paper defects.

[0004] According to a first aspect of the present application, a label paper defect detection method is provided, comprising the following steps:

[0005] S100, for any to-be-detected process in the label paper production process, a plurality of historical monitoring videos corresponding to the to-be-detected process are obtained, and each historical monitoring video is taken as a to-be-identified video; the historical monitoring video corresponds to the operation cycle corresponding to the to-be-detected process.

[0006] S200, for any to-be-identified video, a computer vision model is used to identify the behavior of a plurality of preset to-be-identified objects in the to-be-identified video, and the spatial difference degree, the time difference degree and the sequence difference degree between the to-be-identified video and the preset standard action video are obtained according to the behavior identification result.

[0007] S300, the spatial difference degree, the time difference degree and the sequence difference degree are normalized respectively, and the preset index weight corresponding to the spatial difference, the time difference and the sequence difference is used for weighted sum calculation to obtain a video anomaly score value corresponding to the to-be-identified video.

[0008] S400, the to-be-identified video corresponding to the video anomaly score value greater than the preset score threshold is screened out as a behavior identification abnormal video, so as to determine a plurality of to-be-detected label papers corresponding to each behavior identification abnormal video.

[0009] S500, using an automatic paging device to page a plurality of to-be-detected label papers, and sequentially acquiring a target image of each to-be-detected label paper in the paging process.

[0010] S600, based on a pre-trained deep learning model, performing defect identification on the target image of each to-be-detected label paper to obtain a target defect type of the to-be-detected label paper and to reject the to-be-detected label paper with defects; a plurality of preset defect types are set in the deep learning model.

[0011] According to a second aspect of the present application, a label paper defect detection system is provided, and the system comprises:

[0012] The first acquisition module is configured to, for any to-be-detected process in a label paper production process, acquire a plurality of historical monitoring videos corresponding to the to-be-detected process and take each historical monitoring video as a to-be-identified video; the historical monitoring video corresponds to an operation cycle corresponding to the to-be-detected process.

[0013] The second acquisition module is configured to, for any to-be-identified video, perform behavior identification on a plurality of preset to-be-identified objects in the to-be-identified video based on a computer vision model, and acquire a spatial difference degree, a time difference degree and a sequence difference degree between the to-be-identified video and a preset standard action video according to a behavior identification result.

[0014] The calculation module is configured to perform normalization processing on the spatial difference degree, the time difference degree and the sequence difference degree respectively, and perform weighted sum calculation based on preset index weights corresponding to the spatial difference, the time difference and the sequence difference to obtain a video anomaly score value corresponding to the to-be-identified video.

[0015] The screening module is configured to screen out the to-be-identified video with a video anomaly score value greater than a preset score threshold as a behavior identification abnormal video, so as to determine a plurality of to-be-detected label papers corresponding to each behavior identification abnormal video.

[0016] The third acquisition module is configured to use an automatic paging device to page a plurality of to-be-detected label papers, and sequentially acquire a target image of each to-be-detected label paper in the paging process.

[0017] The identification module is configured to, based on a pre-trained deep learning model, perform defect identification on the target image of each to-be-detected label paper to obtain a target defect type of the to-be-detected label paper and to reject the to-be-detected label paper with defects; a plurality of preset defect types are set in the deep learning model.

[0018] The present application has at least the following beneficial effects:

[0019] The application provides a label paper defect detection method, first, a plurality of historical monitoring videos corresponding to a to-be-detected process are acquired and taken as to-be-recognized videos, behavior recognition is performed on a plurality of preset to-be-recognized objects in the to-be-recognized videos, and spatial difference degrees, time difference degrees and sequence difference degrees from a standard action video are acquired, then a video anomaly score value corresponding to the to-be-recognized videos is calculated by combining the respective preset index weights, the corresponding to-be-detected label papers are found out through the screened behavior recognition anomaly videos, the difference degrees in a plurality of dimensions between the to-be-recognized videos and the standard action video are acquired, the calculated video anomaly score value is more reasonable and reliable, so that the work specification degree of the staff in the to-be-recognized videos can be judged, and the probability of screening the label papers with quality problems is improved, then a deep learning model is used to perform defect recognition on a target image of the to-be-detected label paper, a target defect type is obtained, and the to-be-detected label paper with defects is rejected, the to-be-detected label papers with higher defect probabilities are screened and detected, the sampling inspection quality is improved, the to-be-detected label papers do not need to be detected, and the label paper defect detection efficiency is also improved. BRIEF DESCRIPTION OF DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiment description. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0021] Figure 1 A flow chart of the label paper defect detection method provided by the first embodiment of the present application is shown in the figure.

[0022] Figure 2 A structural schematic diagram of the label paper defect detection system provided by the second embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0023] The technical solutions in the embodiments of the present application will be described clearly and completely in the following with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the present application.

[0024] Embodiment one

[0025] The first embodiment of the present application provides a label paper defect detection method, as shown in the figure, the method comprises the following steps: Figure 1

[0026] ​S100, for any to-be-detected process in a label paper production process, obtaining a plurality of historical monitoring videos corresponding to the to-be-detected process and taking each historical monitoring video as a to-be-identified video; the historical monitoring video corresponds to a work cycle corresponding to the to-be-detected process. For example, the to-be-detected process can be a process of replacing a die, loading raw materials, etc.

[0027] Specifically, the historical monitoring video corresponding to the work cycle corresponding to the to-be-detected process means that the time length of the historical monitoring video is equal to the work cycle corresponding to the to-be-detected process, and the starting point and the ending point of the historical monitoring video correspond to the starting point and the ending point of the work cycle corresponding to the to-be-detected process, respectively. For example, when the to-be-detected process is loading raw materials, the amount of raw materials loaded each time is the same, and the corresponding work cycle is fixed.

[0028] Further, the S100 step further comprises the following steps:

[0029] S101, when the work cycle corresponding to the to-be-detected process is greater than a preset time threshold, calculating the ratio of the work cycle corresponding to the to-be-detected process and the average work time corresponding to the to-be-detected process obtained in advance, and taking the integer part to obtain the frame extraction number k. The person skilled in the art sets the preset time threshold according to the actual demand, for example, 60 seconds.

[0030] It should be noted that the frame extraction number is obtained by the above-mentioned manner to ensure that the extracted image can contain an image of the worker in a working state.

[0031] S102, dividing the historical monitoring video by the frame extraction number k and extracting frames at the division to obtain k key images, and extending the front frame image and the rear frame image of each key image by a preset step length to obtain a key video corresponding to each key image. The person skilled in the art sets the preset step length according to the actual demand, for example, when the key image is at the 20th second and the preset step length is 2 seconds, the obtained key video is from the 18th second to the 22nd second.

[0032] S103, analyzing the action semantic information in each key video, and determining the key video corresponding to the most actions as a target video according to the action semantic information in each key video and the action corresponding relationship in the preset standard action video. In a specific implementation, a computer vision model can be used to identify the action semantic information in the key video.

[0033] S104, according to the total time length of the standard action video and the corresponding time period of the standard action video and the target video, the target video is extended before and after to obtain a to-be-identified video; it can be understood that: the time length of the to-be-identified video is consistent with the time length of the standard action video; in a specific implementation, the starting and ending actions of the to-be-identified video can also be checked according to the starting and ending actions of the standard action video, so as to determine whether to further extend the to-be-identified video.

[0034] When the historical monitoring video is long, the to-be-identified video of the worker in the working state is extracted from the historical monitoring video in the above manner, so that it is not necessary to process all the videos in subsequent processing, the data processing amount is reduced, and at the same time, the action interference in the time period of the non-working state can be prevented, and the reliability of the subsequent video abnormal score is improved.

[0035] S200, for any to-be-identified video, the behavior of a plurality of preset to-be-identified objects in the to-be-identified video is recognized based on a computer vision model, and the spatial difference degree, the time difference degree and the sequence difference degree between the to-be-identified video and the preset standard action video are obtained according to the behavior recognition result.

[0036] Further, the standard action video is constructed by the following steps:

[0037] S10, obtaining a process demonstration flow shot at a preset angle to obtain an initial action video; it can be understood that: the preset angle is consistent with the shooting angle of the historical monitoring video.

[0038] S20, according to each key time node in the initial action video, the initial action video is disassembled into a plurality of discrete action units; the key time node is the node corresponding to the preset key action; it can be understood that: the user labels the key time node.

[0039] S30, labeling the spatial action and time threshold of each discrete action unit, and constructing the standard action video after sorting a plurality of discrete action units. The spatial action includes the action trajectory of the worker and the action trajectory and final position of the tool or object.

[0040] In one specific embodiment, the S200 step includes the following steps:

[0041] S201, extracting the moving track of the preset body part coordinates of the to-be-identified object of the person type from the to-be-identified video, and tracking the tool moving track of the to-be-identified object of the tool type. For example, when replacing the cutter die, the moving track of the wrist, shoulder and other skeleton joint coordinates in the video is extracted by using a pose estimation model; the moving track of the cutter die and the screwing action is tracked by using a target detection model.

[0042] S202, align the time axis of the video to be identified and the standard action video, and match the starting frames of the same action units to obtain standard movement trajectories corresponding to the movement trajectories of the preset body part coordinates and the tool movement trajectories from the standard action video; the preset standard action video includes a plurality of action units.

[0043] S203, for any action unit, according to the movement trajectories of the preset body part coordinates, the tool movement trajectories, and the corresponding standard movement trajectories, the spatial difference value, the time difference value and the sequence difference value corresponding to the action unit are calculated.

[0044] Specifically, the spatial difference value corresponding to the action unit refers to the Euclidean distance of the movement trajectory end point of the same tool in the action unit corresponding to the video to be identified and the standard action video; it can be understood as: obtaining the deviation distance of the final installation position of the tool. In specific implementation, when there are multiple tools, the average Euclidean distance is calculated according to the Euclidean distance of the movement trajectory end point of each tool.

[0045] Preferably, the time difference value corresponding to the action unit refers to the difference value of the time used by the video to be identified and the standard action video to complete the action unit.

[0046] Specifically, the sequence difference value refers to the value obtained after data preprocessing of the sequence difference state of each action in the action unit corresponding to the video to be identified. In specific implementation, ordinal encoding is used to map the sequence to an integer, and the value after preprocessing is obtained according to the change difference value of the sequence. Those skilled in the art know the specific implementation process of preprocessing ordinal data, which will not be described here.

[0047] S204, according to the spatial difference value, the time difference value and the sequence difference value corresponding to each action unit, the spatial difference degree, the time difference degree and the sequence difference degree of the video to be identified and the preset standard action video are calculated; it can be understood that the sum of the spatial difference values corresponding to a plurality of action units is determined as the spatial difference degree, the time difference values corresponding to a plurality of action units are determined as the time difference degree, and the sequence difference values corresponding to a plurality of action units are determined as the sequence difference degree.

[0048] The above, taking the standard action video as a reference, obtains the difference degree of multiple dimensions between the video to be identified and the standard action video, which can judge the work specification degree of the worker in the video to be identified. Since the work specification degree reflects the production quality of the product to a great extent, the above method is helpful to improve the probability of screening out label paper with quality problems.

[0049] S300, normalize the spatial difference degree, the time difference degree and the sequence difference degree respectively, and perform weighted sum calculation based on the preset index weight corresponding to the spatial difference, the time difference and the sequence difference to obtain a video anomaly score value corresponding to the to-be-identified video; it can be understood that: according to the spatial difference degree corresponding to each to-be-identified video, the spatial difference degree is normalized to 0-1, and the time difference degree and the sequence difference degree are the same.

[0050] Specifically, the preset index weight corresponding to the spatial difference, the time difference and the sequence difference is obtained by the following steps:

[0051] S310, according to each preset process in the preset index weight library and the preset index weight table corresponding to each preset process, determine the preset index weight table corresponding to the to-be-detected process.

[0052] S320, according to the preset index weight table corresponding to the to-be-detected process, obtain the preset index weight corresponding to the spatial difference, the time difference and the sequence difference. For example, when installing the cutter die, the spatial difference has a greater impact, that is, whether the position information and angle of the installed cutter die are reliable, when installing the assembly, the sequence plays a key role, and when replacing the glue in the label paper production process, the replacement needs to be completed quickly.

[0053] Because the influence of different indexes on product production is inconsistent in different processes, the index weight corresponding to the process is set according to different processes, so that the calculated video anomaly score value is more reasonable and reliable, and it is beneficial to screen out the to-be-detected label paper with quality problems.

[0054] S400, screen out the to-be-identified video corresponding to the video anomaly score value greater than the preset score threshold value as the behavior recognition abnormal video, so as to determine a plurality of to-be-detected label papers corresponding to each behavior recognition abnormal video. The person skilled in the art can set the preset score threshold value according to the actual needs, which will not be described here.

[0055] Specifically, the determination of a plurality of to-be-detected label papers corresponding to each behavior recognition abnormal video comprises the following steps:

[0056] S401, obtain the monitoring time period and the target process of the historical monitoring video corresponding to each behavior recognition abnormal video.

[0057] S402, according to the product batch corresponding to the target process in the monitoring time period, determine a plurality of to-be-detected label papers corresponding to each behavior recognition abnormal video.

[0058] Since the higher the video anomaly score value is, the lower the motion standard degree of the staff is, and the greater the probability of the label paper with quality problems is, the video with a high video anomaly score value is screened out first, and the corresponding label paper to be detected is found for focused detection, thereby improving the sampling effect, and without detecting all the label papers, the label paper defect detection efficiency is also improved.

[0059] S500, using an automatic paging device to page a plurality of label papers to be detected, and sequentially acquiring a target image of each label paper to be detected in the paging process.

[0060] S600, based on a pre-trained deep learning model, defects of each label paper to be detected are identified, target defect types of the label paper to be detected are obtained, and the label paper to be detected with defects is removed; a plurality of preset defect types are set in the deep learning model. In training, the deep learning model is used to learn and train different defect types and corresponding large numbers of label paper image samples, so that the trained deep learning model can predict the corresponding defect types according to the image samples of the label paper; the specific training process of the deep learning model is known to those skilled in the art, and will not be described here.

[0061] Specifically, the plurality of preset defect types include but are not limited to stains, color differences, missing characters, label paper defects and size deviations.

[0062] The above, by fusing the deep learning model in the automatic paging device, the image of each label paper can be photographed in the process of rapid paging, and the defect type of each label paper is predicted in real time, and the label paper with defects is removed in time, thereby improving the defect detection efficiency and detection quality of the label paper.

[0063] Embodiment two

[0064] Embodiment two of the present application provides a label paper defect detection system, as shown in the figure, the system comprises: Figure 2

[0065] The first acquisition module 100 is used for acquiring a plurality of historical monitoring videos corresponding to a to-be-detected process for any to-be-detected process in a label paper production process, and each historical monitoring video is used as a to-be-identified video; the historical monitoring video corresponds to a work cycle corresponding to the to-be-detected process.

[0066] The second acquisition module 200 is used for, for any to-be-identified video, based on a computer vision model, identifying a plurality of preset to-be-identified objects in the to-be-identified video, and acquiring spatial difference degree, time difference degree and sequence difference degree between the to-be-identified video and the preset standard motion video according to the behavior identification result.

[0067] ​The computing module 300 is configured to normalize the spatial difference degree, the time difference degree and the sequence difference degree respectively, and to calculate a weighted sum based on preset index weights corresponding to the spatial difference, the time difference and the sequence difference, to obtain a video anomaly score value corresponding to the to-be-identified video.

[0068] The screening module 400 is configured to screen out, as a behavior recognition anomaly video, a to-be-identified video corresponding to a video anomaly score value greater than a preset score threshold, to determine a plurality of to-be-detected label papers corresponding to each behavior recognition anomaly video.

[0069] The third obtaining module 500 is configured to perform pagination on the plurality of to-be-detected label papers by using an automatic pagination device, and to sequentially obtain a target image of each to-be-detected label paper in the pagination process.

[0070] The recognition module 600 is configured to perform defect recognition on the target image of each to-be-detected label paper based on a pre-trained deep learning model, to obtain a target defect type of the to-be-detected label paper and to reject the to-be-detected label paper with defects; and the deep learning model is provided with a plurality of preset defect types.

[0071] It should be noted that the information interaction, execution process and the like between the above modules are based on the same concept as the method embodiments of the present application, and the specific functions and the technical effects brought by the same can be referred to the method embodiments part, and will not be repeated here.

[0072] Although some specific embodiments of the present application have been described in detail through examples, those skilled in the art should understand that the above examples are only for illustration, but not for limiting the scope of the present application. Those skilled in the art should also understand that various modifications can be made to the embodiments without departing from the scope and spirit of the present application. The scope of the present application is defined by the appended claims.

Claims

1. A method for detecting defects in label paper, characterized in that, The method includes the following steps: S100: For any process to be inspected in the label paper production process, acquire several segments of historical monitoring video corresponding to the process to be inspected and use each historical monitoring video as a video to be identified; the historical monitoring video corresponds to the work cycle corresponding to the process to be inspected. S200: For any video to be identified, perform behavior recognition on several preset objects in the video based on a computer vision model. Based on the behavior recognition results, obtain the spatial difference, temporal difference, and sequential difference between the video to be identified and the preset standard action video. Step S200 includes the following steps: S201: Extract the movement trajectory of preset body part coordinates of the human-type object to be identified from the video to be identified, and track the tool movement trajectory of the tool-type object to be identified; S202: Align the timelines of the video to be identified and the standard action video, and match the starting frames of the same action unit to obtain the movement trajectory of the preset body part coordinates and the tool movement trajectory from the standard action video. The movement trajectories correspond to standard movement trajectories respectively; the preset standard motion video includes several motion units; S203, for any motion unit, based on the movement trajectory of the preset body part coordinates, the tool movement trajectory, and the corresponding standard movement trajectories, the spatial difference value, temporal difference value, and sequence difference value corresponding to the motion unit are calculated; wherein, the sequence difference value refers to the value obtained after data preprocessing of the sequence difference state of each action in the corresponding motion unit of the video to be recognized; S204, based on the spatial difference value, temporal difference value, and sequence difference value corresponding to each motion unit, the degree of spatial difference, degree of temporal difference, and degree of sequence difference between the video to be recognized and the preset standard motion video are calculated; S300 normalizes the degree of spatial difference, the degree of temporal difference, and the degree of sequential difference, and calculates the video anomaly score value corresponding to the video to be identified by weighted sum based on the preset index weights corresponding to spatial difference, temporal difference, and sequential difference. S400, the videos to be identified that have a corresponding video anomaly score value greater than the preset score threshold are selected as behavior recognition anomaly videos, so as to determine a number of corresponding labels to be detected based on each behavior recognition anomaly video. S500 uses an automatic paging device to paginate several labels to be tested, and sequentially acquires the target image of each label to be tested during the paging process; S600, based on a pre-trained deep learning model, performs defect identification on the target image of each label to be detected, obtains the target defect type of the label to be detected, and removes the defective label to be detected; the deep learning model is set with several preset defect types.

2. The defect detection method for label paper according to claim 1, characterized in that, Step S100 also includes the following steps: S101, when the work cycle corresponding to the process to be inspected is greater than the preset time threshold, calculate the ratio of the work cycle corresponding to the process to be inspected to the average work time corresponding to the process to be inspected obtained in advance, and round down to obtain the number of frame extractions k. S102, the historical surveillance video is divided into average segments according to the number of frame extractions k, and frames are extracted at the segmentation points to obtain k key images. Each key image is expanded with the previous frame image and the next frame image with a preset step size to obtain the key video corresponding to each key image. S103, Analyze the action semantic information in each key video, and determine the key video with the most corresponding actions as the target video based on the action semantic information in each key video and the action correspondence in the preset standard action video. S104. Based on the total duration of the standard action video and the corresponding time period between the standard action video and the target video, the target video is expanded before and after to obtain the video to be identified.

3. The defect detection method for label paper according to claim 1, characterized in that, The correspondence between the historical monitoring video and the work cycle corresponding to the process to be inspected means that the duration of the historical monitoring video is equal to the work cycle corresponding to the process to be inspected, and the start and end points of the historical monitoring video correspond to the start and end points of the work cycle corresponding to the process to be inspected, respectively.

4. The defect detection method for label paper according to claim 1, characterized in that, Create a standard motion video using the following steps: S10: Obtain the process demonstration video shot from a preset angle to get the initial motion video; S20, based on each key time node in the initial motion video, the initial motion video is decomposed into several discrete motion units; the key time node is the node corresponding to the preset key action; S30: Label each discrete action unit with spatial motion and temporal threshold, and construct a standard motion video by sorting several discrete action units.

5. The defect detection method for label paper according to claim 1, characterized in that, In step S300, the preset index weights corresponding to spatial difference, temporal difference, and order difference are obtained through the following steps: S310, Based on each preset process in the preset index weight library and the preset index weight table corresponding to each preset process, determine the preset index weight table corresponding to the process to be tested. S320: Based on the preset index weight table corresponding to the process to be inspected, obtain the preset index weights corresponding to spatial difference, time difference and sequence difference respectively.

6. The defect detection method for label paper according to claim 1, characterized in that, In step S400, determining the corresponding number of labels to be detected based on each behavior of the abnormal video includes the following steps: S401, Obtain the monitoring time period and target process of the historical monitoring video corresponding to each behavior identification abnormal video; S402, based on the product batch corresponding to the target process within the monitoring time period, determine the number of labels to be inspected corresponding to each line of the video that identifies the abnormality.

7. The defect detection method for label paper according to claim 1, characterized in that, The preset defect types include stains, color differences, missing text, damaged labels, and size deviations.

8. A defect detection system for label paper, characterized in that, The system includes: The first acquisition module is used to acquire several segments of historical monitoring videos corresponding to any process to be inspected in the label paper production process, and to use each historical monitoring video as a video to be identified; the historical monitoring videos correspond to the work cycle of the process to be inspected. The second acquisition module is used to perform behavior recognition on several preset objects in any given video to be recognized, based on a computer vision model. Based on the behavior recognition results, it acquires the spatial, temporal, and sequential differences between the video to be recognized and a preset standard action video. The second acquisition module is further used to: extract the movement trajectory of preset body part coordinates of human-type objects from the video to be recognized, and track the tool movement trajectory of tool-type objects; align the timelines of the video to be recognized and the standard action video, and match the starting frames of the same action units to obtain the movement trajectory of the preset body part coordinates and the tool movement trajectory from the standard action video. The movement trajectories correspond to standard movement trajectories; the preset standard motion video includes several motion units; for any motion unit, based on the movement trajectory of the preset body part coordinates, the tool movement trajectory, and the corresponding standard movement trajectories, the spatial difference value, temporal difference value, and sequence difference value corresponding to the motion unit are calculated; wherein, the sequence difference value refers to the value obtained after data preprocessing of the sequence difference state of each action in the corresponding motion unit of the video to be recognized; based on the spatial difference value, temporal difference value, and sequence difference value corresponding to each motion unit, the degree of spatial difference, temporal difference, and sequence difference between the video to be recognized and the preset standard motion video are calculated; The calculation module is used to normalize the degree of spatial difference, the degree of temporal difference, and the degree of sequential difference, and calculate the video anomaly score value corresponding to the video to be identified by weighted sum based on the preset index weights corresponding to the spatial difference, temporal difference, and sequential difference. The filtering module is used to filter out videos to be identified whose corresponding video anomaly score value is greater than a preset score threshold as abnormal behavior recognition videos, so as to determine a number of corresponding labels to be detected based on each abnormal behavior recognition video. The third acquisition module is used to paginate several labels to be tested using an automatic pagination device, and to acquire the target image of each label to be tested sequentially during the pagination process. The identification module is used to identify defects in the target image of each label paper to be detected based on a pre-trained deep learning model, obtain the target defect type of the label paper to be detected, and remove the defective label paper to be detected; the deep learning model is configured with several preset defect types.

Images