A method for online exam cheating identification based on multiple instance learning

Through multi-instance learning method, fragment-level features of video data are extracted and tags are generated. Combined with interpretable feature encoder and attention mechanism, the accuracy and efficiency of cheating behavior recognition in online exams are solved, and efficient cheating behavior recognition is achieved. and positioning.

CN114743137BActive Publication Date: 2024-11-05DALIAN UNIV OF TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210328207.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-03-31
Publication Date
2024-11-05
Estimated Expiration
2042-03-31

AI Technical Summary

Technical Problem

It is difficult for the prior art to accurately identify cheating behaviors in online exams, especially behaviors such as tilting heads and prohibited items. The false alarm rate is high, the detection efficiency is low, and the specific judgment standards are lacking.

Method used

Using a multi-instance learning method, through fragment-level feature extraction and continuous sampling of video data, a multi-instance generator is used to generate fragment-level tags, and input them into the interpretable feature encoder for training, and automatically mark the screen with the attention mechanism. the suspicious location to achieve accurate identification of cheating behavior.

Benefits of technology

It improves the accuracy and detection efficiency of cheating behaviors, provides specific judgment standards, can quickly locate suspicious locations in the space, reduces personnel burden, and improves the efficiency and accuracy of cheating detection.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114743137B_ABST
    Figure CN114743137B_ABST
Patent Text Reader

Abstract

The invention discloses an online examination cheating identification method based on multi-instance learning, and relates to the technical field of examination cheating identification. The invention comprises the following steps: acquiring video information and extracting video features; continuously sampling the extracted video features to determine abnormal videos in the video features; and generating a segment label for each abnormal video using a multi-instance generator for the abnormal videos; the invention can accurately and easily discover cheating behavior through the implementation of the cheating identification method, and has high detection efficiency, specific judgment criteria, high accuracy, and can accurately locate data with little cheating detection, thereby solving the problem of being difficult to collect, and can quickly find suspicious locations in space, and can also solve the detection problem existing in a weak supervision mode, thereby reducing the burden on personnel, improving the efficiency and accuracy of cheating detection and identification, and facilitating personnel use.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of examination cheating identification, and in particular to an online examination cheating identification method based on multi-instance learning. Background Art

[0002] Cheating in an examination refers to the behavior of a candidate taking the examination through improper means, seeking or attempting to seek answers within the scope not permitted by the examination process, when the invigilator is testing the candidate's knowledge and skills through written or oral questions or practical operations, which is contrary to the principles of fairness and justice.

[0003] The existing technology simply calculates whether the facial data of the examinee is greater than a certain threshold. For example, the gaze tracking method is used to determine the deviation angle of the gaze in China Invention Patent Publication No. 112464793. Similarly, China Invention Patent Publication No. 110837784 uses the tilt angle of the head as the main basis for distinguishing cheating, but the setting of the threshold will increase the false alarm rate accordingly. In addition, face recognition technology is introduced to detect whether the examinee has left the seat or collaborated with multiple people, but it cannot further judge the head tilt, prohibited items and other behaviors. In most online exams, it is not easy to detect cheating behaviors simply by relying on a certain human body information. For example, the use of mobile phones and flipping books are mainly concentrated in the movement of the upper body, especially the head. The above-mentioned method for estimating the whole body skeleton posture is too broad, the detection efficiency is low, there is no specific judgment standard, the accuracy is not high, and it is not practical. For this reason, we propose an online exam cheating recognition method based on multi-instance learning to solve the above problems. Summary of the invention

[0004] The purpose of the present invention is to solve the problems in the prior art that the prior art simply calculates whether the facial data of the examinee is greater than a certain threshold, such as: the direction of the line of sight, the tilt angle of the head, etc., with a high false alarm rate; or uses face recognition technology to detect whether the examinee has left the seat, collaborated with multiple people, etc., but cannot further judge behaviors such as tilting the head and carrying prohibited items; and in most online exams, it is not easy to detect cheating behaviors simply by relying on a certain specific human body information, such as using a mobile phone, turning pages, etc., which are mainly concentrated in the movement of the upper body, especially the head. The above-mentioned method for estimating the whole body skeleton posture has too broad conditions, low detection efficiency, no specific judgment criteria, low accuracy, and is not practical. A method for identifying cheating in online exams based on multi-instance learning is proposed.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] An online exam cheating identification method based on multiple instance learning comprises the following steps:

[0007] Get video data and video-level labels;

[0008] Extracting segment-level features of the video data and performing continuous sampling;

[0009] Generate a segment-level label for each abnormal video using a multi-instance generator for the video-level label;

[0010] Inputting the segment label into an interpretable feature encoder to train the interpretable feature encoder;

[0011] Performing feature map extraction on the extracted video features;

[0012] The extracted feature map is input into the trained interpretable feature encoder, and the suspicious positions in the picture are automatically marked using the attention mechanism.

[0013] Preferably, extracting the segment-level features of the video data and performing continuous sampling, the steps are as follows:

[0014] Extracting segment-level features of the video data;

[0015] Continuously dividing the video segment-level features to obtain multiple subsets;

[0016] Preferably, the video-level label is generated using a multi-instance generator to generate a segment-level label for each abnormal video, and the steps are as follows:

[0017] The multi-instance generator predicts instance-level scores for features based on video-level labels;

[0018] Calculating scores of the plurality of subsets after division;

[0019] Calculate the loss function according to the calculation result;

[0020] Calculate the ranking loss according to the loss function to obtain a ranking loss score;

[0021] By introducing sparsity and smoothness constraints in the ranking loss score, we obtain the deep ranking loss;

[0022] Combine the loss function with the deep ranking loss to train a multi-instance generator that can generate segment-level labels;

[0023] Determine abnormal video clips based on the scores;

[0024] A multi-instance generator calculates instance-level labels based on the abnormal fragment scores;

[0025] Preferably, the segment label is input into an interpretable feature encoder to train the interpretable feature encoder, and the training steps of the interpretable feature encoder are as follows:

[0026] Get video clips and corresponding label data;

[0027] Inputting the video clip into an interpretable feature encoder for calculation to obtain a calculation result;

[0028] According to the calculation results, the cross entropy loss function is used to train an interpretable feature encoder with visualization capabilities;

[0029] Preferably, the interpretable feature encoder consists of two branches: a visualization branch and an auxiliary branch. The visualization branch needs to extract feature maps from the penultimate convolutional layer of the feature encoder and the feature maps of the last convolutional layer. At the same time, the feature encoder introduces two submodules consisting of convolutional layers for encoding. The auxiliary branch directly performs global average pooling on the feature maps encoded by the submodules, and predicts scores through activation functions to enhance the visualization capability of the feature encoder.

[0030] An online examination cheating identification device based on multi-instance learning, comprising:

[0031] Video information acquisition module: used to obtain video data and video-level tags;

[0032] Continuous sampling module: used for extracting segment-level features of the video data and performing continuous sampling;

[0033] A segment label generation module: used for generating a segment label for each abnormal video of the multi-instance generator;

[0034] Encoder training module: used for inputting the segment label into the interpretable feature encoder to train the interpretable feature encoder;

[0035] Feature map extraction module: used for extracting feature maps of the extracted video features;

[0036] Position standard module: used to input the extracted feature map into the trained interpretable feature encoder and automatically mark suspicious positions in the picture using the attention mechanism.

[0037] Preferably, the fragment tag generating module includes:

[0038] Score prediction unit: A multi-instance generator predicts instance-level scores for features based on video-level labels;

[0039] Score division unit: calculating the scores of the plurality of subsets after division;

[0040] Loss function calculation unit: calculates the loss function according to the calculation result;

[0041] A ranking loss score calculation unit: calculates the ranking loss according to the loss function to obtain a ranking loss score;

[0042] Deep ranking loss derivation unit: By introducing sparsity and smoothness constraints in the ranking loss score, the deep ranking loss is obtained;

[0043] Loss function combination unit: Combine the loss function with the deep ranking loss to train a multi-instance generator that can generate segment-level labels;

[0044] Abnormal segment determination unit: determines abnormal video segments according to the scores;

[0045] Label generation unit: a multi-instance generator calculates instance-level labels according to the abnormal fragment scores;

[0046] Preferably, the encoder training module comprises:

[0047] Data acquisition unit: acquires video clips and corresponding label data;

[0048] A calculation unit: used for inputting the video clip into an interpretable feature encoder for calculation to obtain a calculation result;

[0049] Training unit: Based on the calculation results, the cross entropy loss function is used to train an interpretable feature encoder with visualization capabilities;

[0050] Compared with the prior art, the beneficial effects of the present invention are as follows: the present invention can accurately and easily detect cheating behavior through the implementation of the cheating identification method, and has high detection efficiency, has specific judgment criteria, has high accuracy, and can accurately locate data with little cheating detection, solving the problem of difficult collection, and can quickly find suspicious locations in space. It can also solve the detection problem under weak supervision, reduce the burden on personnel, improve the efficiency and accuracy of cheating detection and identification, and facilitate personnel to use. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 A schematic diagram of the processing process of an online examination cheating identification method based on multi-instance learning proposed by the present invention;

[0052] Figure 2 This is a schematic diagram of the overall process of an online exam cheating identification method based on multi-instance learning proposed by the present invention;

[0053] Figure 3 A schematic flow chart of the steps of continuous sampling of an online examination cheating identification method based on multi-instance learning proposed by the present invention;

[0054] Figure 4A schematic flow chart of the steps of generating a segment label for an online examination cheating identification method based on multiple instance learning proposed by the present invention;

[0055] Figure 5 A flowchart of the training steps of an interpretable feature encoder for an online exam cheating identification method based on multiple instance learning proposed by the present invention. DETAILED DESCRIPTION

[0056] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0057] Reference Figure 1-5 , an online exam cheating identification method based on multiple instance learning, comprising the following steps:

[0058] S1: Obtain video data and video-level labels;

[0059] S2: extracting the segment-level features of the video data and performing continuous sampling, the steps are as follows:

[0060] S201: extracting segment-level features of the video data, N is the total number of features.

[0061] S202: Continuously divide the video segment level features to obtain multiple subsets. L represents the number of subsets in each video, T represents the number of segments contained in each subset; U is the subset;

[0062] S3: Generate a segment-level label for each abnormal video using a multi-instance generator for the video-level label, the steps are as follows:

[0063] S301: Multi-instance generator predicts instance-level scores for features based on video-level labels

[0064] S302: Calculate the scores of the plurality of subsets after division, using the following formula: In the formula, S l Score the subset;

[0065] S303: Calculate the loss function according to the calculation result. The calculation formula is as follows:

[0066]

[0067] Among them, max is the maximum value of all subsets of each video; and is the prediction score for normal and abnormal subsets.

[0068] S304: Calculate the ranking loss according to the loss function to obtain a ranking loss score. The calculation formula is as follows:

[0069]

[0070] In this formula, a video V is regarded as a packet, and the video is split into segments v i As an example; negative bag is marked as It means that it does not contain abnormal instances, Y = 0; while the positive package is marked as It contains at least one abnormal instance, Y = 1; therefore, video anomaly detection can be modeled as instance detection in the MIL problem, that is, searching for abnormal instances v in the positive bag i And assign anomaly scores, so that anomalies can be detected from weak monitoring signals;

[0071] In this paper, after given positive and negative packets, for B a and B n Extract features from each segment and The feature information is then fed into the multi-instance generator to estimate the anomaly score for each segment; the anomaly score is normalized and used as the segment-level label of the corresponding video. And put it into the interpretable encoder for training to locate anomalies in the spatial dimension;

[0072] S305: By introducing sparsity and smoothness constraints in the ranking loss score, the deep ranking loss is obtained, and the formula is as follows:

[0073]

[0074] In the formula, formula 1 is used as a smoothing term, and formula 2 is a sparse term; 1 With λ 2 is a hyperparameter used to balance the ranking loss;

[0075] S306: combining the loss function with the deep ranking loss to train a multi-instance generator capable of generating segment-level labels;

[0076] S307: Determine abnormal video segments according to the scores;

[0077] S308: The multi-instance generator calculates the instance-level label according to the abnormal fragment score, and the formula is as follows:

[0078]

[0079] In the formula, In [0,1], as the fragment label; is the instance-level score corresponding to the abnormal fragment; Represents a collection of anomaly instance-level scores.

[0080] Finally, the generated label data is combined with the corresponding video clip as the input of the feature encoder;

[0081] S4: Input the fragment label into the interpretable feature encoder, and train the interpretable feature encoder. The interpretable feature encoder consists of two branches: a visualization branch and an auxiliary branch. The visualization branch needs to extract the feature map from the penultimate convolutional layer of the feature encoder and the feature map of the last convolutional layer; at the same time, the feature encoder introduces two sub-modules composed of convolutional layers for encoding; the auxiliary branch directly performs global average pooling on the feature map encoded by the sub-module, and predicts the score through the activation function to enhance the visualization ability of the feature encoder. The training steps of the interpretable feature encoder are as follows:

[0082] S401: Obtain video clips and corresponding tag data;

[0083] S402: Input the video clip into an interpretable feature encoder for calculation to obtain a calculation result, and the calculation formula is as follows:

[0084]

[0085] In the formula, is element-wise multiplication; A 3 Finally, global average pooling is applied to the fully connected layer to predict the anomaly score; A 1 is the feature map of the last convolutional layer; A 2 Extract feature maps for the penultimate convolutional layer; M 1 It is a submodule composed of convolutional layers; in the auxiliary branch, the feature map A 2 Through the submodule M 2 Then global average pooling and channel-level average pooling are performed for compression, and the auxiliary score of each category (normal and abnormal) is obtained through the softmax activation function;

[0086] S403: Based on the calculation results, a cross entropy loss function is used to train an interpretable feature encoder with visualization capability;

[0087] S5: extracting a feature map from the extracted video features;

[0088] S6: Input the extracted feature map into the trained interpretable feature encoder, and use the attention mechanism to automatically mark suspicious locations in the picture.

[0089] An online examination cheating identification device based on multi-instance learning, comprising:

[0090] Video information acquisition module: used to obtain video data and video-level tags;

[0091] Continuous sampling module: used to extract the segment-level features of the video data and perform continuous sampling, the steps are as follows:

[0092] extracting segment-level features of the video data, N is the total number of features;

[0093] Continuously dividing the video clip level features to obtain multiple subsets, L represents the number of subsets in each video, T represents the number of segments contained in each subset; U is the subset;

[0094] A segment label generation module: used for generating a segment label for each abnormal video of the multi-instance generator;

[0095] Encoder training module: used to input the segment label into the interpretable feature encoder and train the interpretable feature encoder. The interpretable feature encoder consists of two branches: a visualization branch and an auxiliary branch. The visualization branch needs to extract the feature map from the penultimate convolutional layer of the feature encoder and the feature map of the last convolutional layer. At the same time, the feature encoder introduces two sub-modules composed of convolutional layers for encoding. The auxiliary branch directly performs global average pooling on the feature map encoded by the sub-module, and predicts the score through the activation function to enhance the visualization ability of the feature encoder.

[0096] Feature map extraction module: used for extracting feature maps of the extracted video features;

[0097] Position standard module: used to input the extracted feature map into the trained interpretable feature encoder and automatically mark suspicious positions in the picture using the attention mechanism.

[0098] Among them, the fragment tag generation module includes:

[0099] Score Prediction Unit: A multi-instance generator predicts instance-level scores for features based on video-level labels

[0100] Score division unit: calculates the scores of the plurality of subsets after division, the formula is as follows: In the formula, S l Score the subset;

[0101] Loss function calculation unit: calculates the loss function according to the calculation result, and the calculation formula is as follows:

[0102]

[0103] Among them, max is the maximum value of all subsets of each video; and is the prediction score for normal and abnormal subsets.

[0104] Ranking loss score calculation unit: calculates the ranking loss according to the loss function to obtain the ranking loss score. The calculation formula is as follows:

[0105]

[0106] In this formula, a video V is regarded as a packet, and the video is split into segments v i As an example; negative bag is marked as It means that it does not contain abnormal instances, Y = 0; while the positive package is marked as It contains at least one abnormal instance, Y = 1; therefore, video anomaly detection can be modeled as instance detection in the MIL problem, that is, searching for abnormal instances v in the positive bag i And assign anomaly scores, so that anomalies can be detected from weak monitoring signals;

[0107] In this paper, after given positive and negative packets, for B a and B n Extract features from each segment and The feature information is then fed into the multi-instance generator to estimate the anomaly score for each segment; the anomaly score is normalized and used as the segment-level label of the corresponding video. And put it into the interpretable encoder for training to locate anomalies in the spatial dimension;

[0108] Deep ranking loss derivation unit: By introducing sparsity and smoothness constraints in the ranking loss score, the deep ranking loss is obtained as follows:

[0109]

[0110] In the formula, formula 1 is used as a smoothing term, and formula 2 is a sparse term; 1 With λ 2 is a hyperparameter used to balance the ranking loss;

[0111] Loss function combination unit: Combine the loss function with the deep ranking loss to train a multi-instance generator that can generate segment-level labels;

[0112] Abnormal segment determination unit: determines abnormal video segments according to the scores;

[0113] Label generation unit: The multi-instance generator calculates the instance-level label according to the abnormal fragment score, and the formula is as follows:

[0114]

[0115] In the formula, In [0,1], as the fragment label; is the instance-level score corresponding to the abnormal fragment; Represents a collection of anomaly instance-level scores.

[0116] Finally, the generated label data is combined with the corresponding video clip as the input of the feature encoder;

[0117] Among them, the encoder training module includes:

[0118] Data acquisition unit: acquires video clips and corresponding label data;

[0119] Calculation unit: input the video clip into the interpretable feature encoder for calculation to obtain a calculation result. The calculation formula is as follows:

[0120]

[0121] In the formula, is element-wise multiplication; A 3 Finally, global average pooling is applied to the fully connected layer to predict the anomaly score; A 1 is the feature map of the last convolutional layer; A 2 Extract feature maps for the penultimate convolutional layer; M 1 It is a submodule composed of convolutional layers; in the auxiliary branch, the feature map A 2 Through the submodule M 2 Then global average pooling and channel-level average pooling are performed for compression, and the auxiliary score of each category (normal and abnormal) is obtained through the softmax activation function;

[0122] Training unit: Based on the calculation results, the cross entropy loss function is used to train an interpretable feature encoder with visualization capabilities;

[0123] An intelligent computer device comprises a memory and a processor, wherein the memory stores computer-readable instructions, and when the processor executes the computer-readable instructions, the steps of a cheating identification method based on video anomaly detection are implemented.

[0124] A computer-readable storage medium stores computer-readable instructions, and when the computer-readable instructions are executed by a processor, the steps of a cheating identification method based on video anomaly detection are implemented.

[0125] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for identifying cheating in online exams based on multiple instance learning, characterized in that: The steps include: Step 1: Obtain video data and video-level labels; Step 2: extract the segment-level features of the video data and perform continuous sampling, the steps are as follows: extracting segment-level features of the video data, N is the total number of features; Continuously dividing the segment-level features of the video data to obtain multiple subsets, L represents the number of subsets in each video, T represents the number of segments contained in each subset; U is the subset; Step 3: Generate a segment-level label for each abnormal video using a multi-instance generator for the video-level label, the steps are as follows: Multi-instance generator predicts instance-level scores for features based on video-level labels The scores of the multiple subsets after division are calculated using the following formula: In the formula, S l Score the subset; The loss function is calculated based on the calculation results, and the calculation formula is as follows: Among them, max is the maximum value of all subsets of each video; and Prediction scores for normal and abnormal subsets; The ranking loss is calculated according to the loss function to obtain the ranking loss score. The calculation formula is as follows: In this formula, a video V is regarded as a packet, and the video is split into segments v i As an example; negative bag is marked as It means that it does not contain abnormal instances, Y = 0; while the positive package is marked as It contains at least one abnormal instance, Y = 1; therefore, video anomaly detection can be modeled as instance detection in the MIL problem, that is, searching for abnormal instances v in the positive bag i And assign anomaly scores, so that anomalies can be detected from weak monitoring signals; After given the positive and negative packets, for B a and B n Extract features from each segment and The feature information is then fed into the multi-instance generator to estimate the anomaly score for each segment; the anomaly score is normalized and used as the segment-level label of the corresponding video. And put it into the interpretable encoder for training to locate anomalies in the spatial dimension; By introducing sparsity and smoothness constraints in the ranking loss score, the deep ranking loss is obtained, which is formulated as follows: Wherein, formula 1 is a smoothing term, formula 2 is a sparse term; λ1 and λ2 are hyperparameters used to balance the ranking loss; Combine the loss function with the deep ranking loss to train a multi-instance generator that can generate segment-level labels; Determine abnormal video clips based on the scores; The multi-instance generator calculates the instance-level label based on the abnormal video segment score, and the formula is as follows: In the formula, In [0,1], as the fragment label; is the instance-level score corresponding to the abnormal fragment; represents the set of abnormal instance-level scores; Finally, the generated label data is combined with the corresponding video clip as the input of the original feature encoder; Step 4: Input the segment label into the interpretable feature encoder and train the interpretable feature encoder. The interpretable feature encoder consists of two branches: a visualization branch and an auxiliary branch. The visualization branch needs to extract the feature map from the penultimate convolutional layer of the original feature encoder and the feature map of the last convolutional layer; at the same time, the original feature encoder introduces two sub-modules composed of convolutional layers for encoding; the auxiliary branch directly performs global average pooling on the feature map encoded by the sub-module, and predicts the score through the activation function to enhance the visualization ability of the feature encoder; The training steps of the interpretable feature encoder are as follows: Get video clips and corresponding label data; The video clip is input into the interpretable feature encoder for calculation to obtain a calculation result, and the calculation formula is as follows: A3=A1+M1(A2)°A1 Where, ° is element-wise multiplication; A3 is finally applied to the fully connected layer through global average pooling to predict the anomaly score; A1 is the feature map of the last convolutional layer; A2 is the feature map extracted by the penultimate convolutional layer; M1 is the submodule composed of the convolutional layer; in the auxiliary branch, feature map A2 is compressed by global average pooling and channel-level average pooling after passing through submodule M2, and the auxiliary scores of normal and abnormal are obtained through the softmax activation function; According to the calculation results, the cross entropy loss function is used to train an interpretable feature encoder with visualization capabilities; Step 5: extracting feature graphs from the extracted segment-level features of the video data; Step 6: Input the extracted feature map into the trained interpretable feature encoder, and use the attention mechanism to automatically mark suspicious locations in the picture.

2. An online examination cheating identification device based on multiple instance learning, using the method of claim 1, characterized in that: include: Video information acquisition module: used to obtain video data and video-level tags; Continuous sampling module: used for extracting segment-level features of the video data and performing continuous sampling; Segment label generation module: Uses a multi-instance generator to generate segment labels for each abnormal video; Encoder training module: used for inputting the segment label into the interpretable feature encoder to train the interpretable feature encoder; Feature map extraction module: used for extracting feature maps from the segment-level features of the extracted video data; Position standard module: used to input the extracted feature map into the trained interpretable feature encoder, and automatically mark the suspicious positions in the picture using the attention mechanism.

3. According to claim 2, the online examination cheating identification device based on multi-instance learning is characterized in that: The fragment tag generation module includes: Score prediction unit: A multi-instance generator predicts instance-level scores for features based on video-level labels; Score division unit: calculating the scores of the plurality of subsets after division; Loss function calculation unit: calculates the loss function according to the calculation result; A ranking loss score calculation unit: calculates the ranking loss according to the loss function to obtain a ranking loss score; Deep ranking loss derivation unit: By introducing sparsity and smoothness constraints in the ranking loss score, the deep ranking loss is obtained; Loss function combination unit: Combine the loss function with the deep ranking loss to train a multi-instance generator that can generate segment-level labels; Abnormal segment determination unit: determines abnormal video segments according to the scores; Label generation unit: A multi-instance generator calculates instance-level labels based on the abnormal video segment scores.

4. The online examination cheating identification device based on multi-instance learning according to claim 2 is characterized in that: The encoder training module includes: Data acquisition unit: acquires video clips and corresponding label data; A calculation unit: used for inputting the video clip into an interpretable feature encoder for calculation to obtain a calculation result; Training unit: Based on the calculation results, the cross entropy loss function is used to train an interpretable feature encoder with visualization capabilities.

5. An intelligent computer device, characterized in that: The system comprises a memory and a processor, wherein the memory stores computer-readable instructions, and the processor implements the steps of the cheating identification method based on video anomaly detection as claimed in claim 1 when executing the computer-readable instructions.

6. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the steps of the cheating identification method based on video anomaly detection according to claim 1 are implemented.