Semi-supervised action quality evaluation method and system based on group comparative learning

Abstract

The invention discloses a semi-supervised motion quality assessment method and system based on group contrast learning, and the method comprises the following steps: S1, video data preprocessing, including target detection and feature extraction in a video; s2, based on a teacher-student network, generating a sub-action pseudo tag at a teacher branch; s3, based on a teacher-student network, designing group comparative learning, and realizing self-supervised sub-action analysis of the motion sequence; s4, on the basis of self-supervised sub-action analysis, motion sequence fine-grained feature expression is extracted; and S5, based on distillation learning, performing semi-supervised action quality evaluation. According to the method, semi-supervised action quality evaluation only needing a small amount of annotated data and a large amount of unannotated data is realized, and the dependence on manual annotation is reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of motion quality assessment, and in particular relates to a semi-supervised motion quality assessment method and system based on group contrastive learning. Background Art

[0002] The Action Quality Assessment (AQA) task aims to design an intelligent scoring system that can automatically and objectively evaluate scores for competitive sports and various movements requiring standardization. Action quality assessment has a wide range of applications, including surgical skill rating and medical rehabilitation training in healthcare, posture correction training systems for athletes in sports, AI refereeing systems in competitions, and operational compliance analysis and risk behavior monitoring in industrial production. The development of artificial intelligence (AI) technology is further promoting the widespread application of action quality assessment technology in a wider range of scenarios.

[0003] When evaluating the technical aspects of competitive sports and standardized human motion sequences, motion quality assessment requires not only a general evaluation of the entire sequence, as these sequences often include diverse variations, but also a focus on whether each key technical movement within the sequence meets standard requirements. Therefore, improving the performance of motion quality assessment primarily involves fine-grained analysis of the sub-movements within the sequence. Another challenge facing motion quality assessment is sample labeling. Manually collecting and labeling data samples requires domain-specific knowledge and expertise, making supervised data collection expensive and relatively scarce. Summary of the Invention

[0004] The main purpose of the present invention is to propose a semi-supervised action quality assessment method and system based on group contrastive learning, which can realize semi-supervised action quality assessment with only a small amount of labeled data and a large amount of unlabeled data, reducing the dependence on manual labeling.

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

[0006] The semi-supervised action quality assessment method based on group contrastive learning includes the following steps:

[0007] S1, video data preprocessing, including target detection and feature extraction in the video;

[0008] S2, based on the teacher-student network, generate sub-action pseudo labels in the teacher branch;

[0009] S3. Based on the teacher-student network, we design group contrastive learning to achieve self-supervised sub-action parsing of motion sequences;

[0010] S4, extracting fine-grained feature expressions of motion sequences based on self-supervised sub-action parsing;

[0011] S5. Perform semi-supervised action quality evaluation based on distillation learning.

[0012] The present invention also includes a semi-supervised action quality assessment system based on group contrastive learning, which adopts the semi-supervised action quality assessment method provided by the present invention. The system includes a preprocessing module, a pseudo-label generation module, a self-supervised sub-action parsing module, a motion sequence fine-grained feature expression extraction module, and a semi-supervised action quality assessment module;

[0013] Preprocessing module, used for video data preprocessing, including target detection and feature extraction in the video;

[0014] The pseudo-label generation module generates pseudo-labels for sub-actions in the teacher branch based on the teacher-student network;

[0015] The self-supervised sub-action parsing module is based on a teacher-student network and designs group contrastive learning to achieve self-supervised sub-action parsing of motion sequences;

[0016] The motion sequence fine-grained feature expression extraction module extracts the fine-grained feature expression of the motion sequence based on the self-supervised sub-action parsing module;

[0017] The semi-supervised action quality assessment module performs semi-supervised action quality assessment based on distillation learning.

[0018] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0019] 1. The present invention adopts self-supervised sub-action parsing based on a teacher-student network; first, the teacher branch network performs human body detection on the motion sequence video frames to obtain the area centered on the athlete; secondly, the area centered on the athlete is extracted, and the extracted features are clustered to form a sub-action subset; high-confidence high-quality pseudo labels are generated, which are used to understand the high-level semantics and internal time dependencies of the sub-actions; finally, the student branch network performs feature extraction on the motion sequence video frames, and clusters the extracted features to form sub-action subsets; then, based on the subsets in each branch, the sub-action sequence is parsed; the action subsets of the teacher branch network provide supervision information in the form of pseudo-labels during the sub-action parsing process, guiding the student branch to obtain sub-action semantic-guided features through consistency regularization learning. In this way, the high-level semantics of the sub-action sequence can be effectively captured, and the teacher-student branch self-supervised sub-action sequence parsing can be realized.

[0020] 2. The present invention adopts a group contrast learning method. By grouping sub-actions with similar semantics in two branches, contrast learning is constructed between sub-action groups, which brings sub-action sequence expressions with the same semantics closer and sub-action sequence expressions with different semantics farther apart; sub-action groups are aggregated by sub-action sequences with the same pseudo-label; the feature mean of the sub-action group is calculated, and the feature mean is used to represent the same sub-action, and a pseudo-label is given. This processing method is beneficial to maintaining semantic consistency between sub-actions of the same category.

[0021] 3. The present invention adopts a distillation learning algorithm to construct two parallel branches of supervised learning and unsupervised learning, which are used to learn consistent semantic representations of labeled and unlabeled samples. It is applied to semi-supervised action quality assessment, realizing semi-supervised action quality assessment with only a small amount of labeled data and a large amount of unlabeled data, reducing dependence on manual labeling. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 is a flow chart of the method of the present invention;

[0023] Figure 2 This is a schematic diagram of the design of group contrastive learning to achieve self-supervised sub-action parsing of motion sequences;

[0024] Figure 3 Schematic diagram of semi-supervised motion quality assessment based on distillation learning. DETAILED DESCRIPTION

[0025] Explanation of related terms:

[0026] I3D network, Inflated 3D ConvNet, that is, expanded convolutional network, expands the 2D model into a 3D model by converting the convolution kernel of the 2D network into a 3D convolution kernel and the 2D pooling layer into a 3D pooling layer, while keeping other structures unchanged. The 2D model can use existing networks such as VGG and ResNet.

[0027] The present invention will be described in further detail below with reference to the embodiments and drawings, but the embodiments of the present invention are not limited thereto.

[0028] Example

[0029] like Figure 1 、 Figure 2 as well as Figure 3 As shown, the present invention, a semi-supervised action quality assessment method based on group contrastive learning, includes the following steps:

[0030] S1. Video data preprocessing, including object detection and feature extraction in the video; specifically, including:

[0031] Prepare a training set, including a set of labeled samples and the unlabeled sample set

[0032] Among them, N l and N u Represents the number of labeled and unlabeled samples, y i Represents Corresponding real scoring data; X l and X u Used for subsequent semi-supervised action quality evaluation;

[0033] Construct a teacher-student network, use a pre-trained target detector in the teacher branch, and l Detect human motion areas in video frames and extract the detected human motion areas to form a new sample set

[0034] X u and X a This is then fed into a teacher-student network for training self-supervised sub-action parsing.

[0035] S2, based on the teacher-student network, generate sub-action pseudo labels in the teacher branch;

[0036] The present invention uses self-supervised learning to realize the positioning and segmentation of sub-actions in motion video sequences. In the teacher-student network, the teacher branch uses the target detector to extract the motion area with concentrated information in the video through human target detection based on the initial video sample, and uses the I3D network set X a The teacher branch learns effective feature expressions that are not affected by scene changes based on the sample features, clusters the effective feature expressions, and gives sub-action pseudo labels based on the clustering results, which subsequently guide the student branch in sub-action analysis.

[0037] Generate pseudo labels for sub-actions in the teacher branch, specifically:

[0038] Using the feature learning network, the attention mechanism is used to perform regional weighting on the video frame features to obtain representative feature expressions; in the teacher branch, the attention model is trained on the set X a The target area of the sample video frame after target detection is weighted, and the attention model after transfer training performs attention weighting on the video frame in the student branch to highlight the key motion information in the video frame;

[0039] The attention weight calculation is as follows:

[0040] The motion sequence features F extracted by the I3D network are w(t) Input feature learning network, w∈{u,a}; perform global average pooling on each channel feature and concatenate the pooled features to obtain a vector of dimension H×W×1;

[0041] Perform convolution and Sigmoid activation operations on the pooled features to calculate the spatial attention weight, attention weight and video feature F of the video frame w (t) Perform multiplication operation to obtain the information-weighted features Here, t represents the sequence time. This feature strengthens the areas most related to motion and weakens the environmental areas in the video frame that have little to do with motion. It enhances the discriminative features required in the sub-action parsing process and helps generate representative pseudo-labels.

[0042] Sub-action pseudo labels are generated by clustering the features, specifically:

[0043] Features after weighted attention on the teacher branch Perform K-means clustering to obtain M cluster centers. Each cluster center corresponds to a pseudo label, and M sub-action pseudo labels are obtained.

[0044] The process of selecting representative feature samples for M clusters adopts a greedy algorithm, specifically:

[0045] Attention-weighted features And M cluster groups, select the significant samples representing each cluster from the M cluster groups, which are defined as the representative sample set representing the cluster center, and use S = {S m ,m=1,…,M} represents;

[0046] The process of screening potential representative samples uses the following formula as the objective function and adopts a greedy algorithm to minimize the objective function:

[0047]

[0048] For the mth cluster center, select the first K samples that minimize the objective function Z(S) from the cluster sample group as the representative samples representing the mth cluster center, and store the K samples in the subset s m , is used as the representative sample of the mth sub-action, and the set of M sub-action sample subsets is S = {s m ,m=1,…,M};

[0049] Encode M cluster centers, the mth sub-action set s m Assign pseudo label p m , the pseudo label set of M sub-actions is expressed as P = {p m,m=1,…,M}; these pseudo labels serve as supervisory information in the sub-action parsing process to achieve self-supervised sub-action parsing.

[0050] S3. Based on the teacher-student network, group contrastive learning is designed to achieve self-supervised sub-action parsing of motion sequences; Figure 2 As shown, specifically including:

[0051] A sub-action classifier is constructed using a multi-layer perceptron and Softmax activation function. The sub-action sample subset S and pseudo-label set P selected from the teacher branch are used as training data for classifier training.

[0052] The classifier is used to classify the sub-actions frame by frame in the video sequences of the teacher-student branches. The feature of each video frame is represented as The sub-action parsing process will feature Mapped into a probability vector, denoted as A={a m ,m=1,…,M}, represents the predicted value of the probability of the sub-action occurring in the video sequence, where a m The predicted probability corresponding to the mth sub-action;

[0053] To ensure accurate prediction, self-supervised learning is used to keep the sub-action pseudo labels and sub-action parsing prediction values a m consistency between them; set a cross entropy loss function L c , using the trained classifier to classify features Perform sub-action prediction and obtain the predicted value Then use the sub-action pseudo label p m As a supervisory signal, by optimizing the cross entropy loss function L c , so that self-supervised learning maintains semantic consistency in feature learning and sub-action parsing.

[0054]

[0055] The goal of self-supervised sub-action parsing is to obtain sub-action sequence classification and its temporal dependencies. Given the feature representation of the motion sequence after attention weighting in the teacher and student branches, In order to obtain representative feature representations and accurately classify and segment sub-actions in motion sequences, this paper designs a group contrast learning method that minimizes the distance between sub-action features with the same semantics and maximizes the distance between sub-actions with different semantics, thereby learning fine-grained feature representations guided by sub-action decomposition. To this end, feature learning uses a multi-layer perceptron module to further refine feature representations, and uses a sub-action classifier to classify feature representations. Mapped to the sub-action category probability distribution vector h a (t),h u(t);

[0056] The sub-action category probability distribution vector is further assigned a pseudo-label p corresponding to the category with the maximum activation value and high semantic similarity. Sub-actions with the same pseudo-label are grouped together. The process of grouping motion sequences into sub-actions is shown in the following formula:

[0057]

[0058] Among them, g(h w ) represents the calculation of h for a sub-action sequence w The average logical value of ,w∈{u,a}, represents the calculation of the two branches,w∈{u,a}; It is an indicator function, which takes the value 1 when the condition is met, T B Indicates the number of sequences in a batch B.

[0059] If two sub-action sequences in the teacher-student network branch correspond to the same sub-action pseudo-label at the same time, then the two sub-action groups become a positive sample pair. Sub-action sequences from the same video group but classified into different pseudo-label groups (such as pseudo-labels p, q) are considered as negative sample pairs.

[0060] The present invention sets up group contrast learning to make the teacher-student network branches close the feature expressions of sub-actions of the same category and distance the feature expressions of different categories. The loss function of sub-action group contrast learning is defined as:

[0061]

[0062] where D(·) denotes the distance calculation, τ is the temperature hyperparameter, and q,p∈P denotes different pseudo labels.

[0063] In this way, the method of the present invention distinguishes different sub-actions that are time-related, ensures that the sub-actions are classified from the action sequence, and locates the time correspondence of the sub-actions from the sequence based on the classification of the video frames.

[0064] S4, extracting fine-grained feature expressions of motion sequences based on self-supervised sub-action parsing;

[0065] S5. Based on distillation learning, semi-supervised action quality evaluation is performed; e.g. Figure 3 As shown, specifically including:

[0066] Two network branches are set up to train supervised and unsupervised feature learning respectively; the supervised branch uses I3D network to extract action features and evaluate the action quality to obtain evaluation scores. These action features and evaluation score labels are used to train the supervised score regression model R for action quality evaluation. l(·), and then provide reference regression model parameters for unsupervised branch model training through distillation learning; the supervised regression loss function is defined as:

[0067]

[0068] Among them, y i Represents the true score label corresponding to sample i;

[0069] In training the unsupervised branch of the score regression model R u (·), the unlabeled data samples are first fed into the pre-trained regressor of the supervised branch to obtain the pseudo-labels of the unsupervised branch. By minimizing the difference between the prediction results of the unsupervised branch and the pseudo-labels generated by the supervised branch, the unsupervised branch is encouraged to learn features and score regressors for accurate prediction of action scores. This process is called consistency regularization, which encourages the unsupervised branch to keep its predictions on unlabeled data consistent with the pseudo-labels generated by the regressor of the supervised branch.

[0070] The semi-supervised motion quality assessment score regression loss function is conceptualized as follows:

[0071]

[0072] Among them, τ is the preset threshold; Represents the indicator function, and the result is 1 when the maximum predicted value exceeds τ, otherwise it is 0. This semi-supervised motion quality assessment score regression loss function aims to use the pre-trained supervised branch to generate high-quality pseudo labels to guide the learning process of the unsupervised branch. By minimizing the difference between the unsupervised branch prediction and the pseudo labels, the unsupervised branch is encouraged to learn features that are semantically consistent with the supervised branch and accurately express the action, thereby effectively utilizing labeled and unlabeled data for semi-supervised motion quality assessment. The core idea of this loss function is to help the unsupervised branch gradually improve its performance on unlabeled samples through the pseudo labels generated by the supervised branch, while ensuring that it maintains good performance on labeled samples.

[0073] In this embodiment, the overall training loss of the semi-supervised action quality assessment method is:

[0074] L=L reg +λ1L sem +λ2L c +λ3L gc

[0075] Among them, L reg is the supervised regression loss for supervised quality assessment, L sem Regularization loss for constrained pseudo-label consistency learning, L c is the self-supervised cross entropy loss, Lgc is the sub-action group contrastive loss, and λ1, λ2, and λ3 are weight parameters.

[0076] The following examples are used to experimentally verify the trained motion quality assessment model in this embodiment. In this experimental verification, two video datasets in sports, MTL-AQA and Rhythmic Gymnastics, are used, and five comparison methods are selected for comparative experiments to statistically evaluate the accuracy. The four comparison methods are Pseudo-labels (Pseudo-label based semi-supervised estimation), VAT (Virtual Adversarial Training), S 4 L (Self-supervised semi-supervised learning), S 4 AQA (semi-supervised AQA).

[0077] Table 1 shows a comparison of the score prediction accuracy of the semi-supervised action quality assessment method in this embodiment and the comparison method on the sports video dataset MTL-AQA.

[0078]

[0079] Table 1

[0080] Table 2 shows a comparison of the score prediction accuracy of the method of this embodiment and the comparative method on the sports video dataset Rhythmic Gymnastics.

[0081]

[0082] Table 2

[0083] According to Table 1, with S 4 Compared to AQA, the semi-supervised motion quality assessment method implemented in this embodiment on the MTL-AQA dataset using 10% and 40% labeled samples achieved performance improvements of 5.3% and 5.5%, respectively. These results demonstrate that the teacher-student branch self-supervised sub-action parsing can effectively improve the overall score prediction accuracy of the algorithm model.

[0084] Comparing the data in Table 2, it can be seen that the method in this embodiment is superior to the most advanced semi-supervised method S 4For AQA, using only 40% supervised information, the method in this example improves accuracy by up to 5.1%. This paper provides a semi-supervised framework that captures fine-grained sub-action information in a self-supervised manner and uses it to accurately assess motion quality. This improvement is achieved by effectively leveraging unlabeled video samples to learn discriminative feature representations and temporal relationships between action sequences.

[0085] In another embodiment, a semi-supervised action quality assessment system based on group contrastive learning is also provided. The system adopts the semi-supervised action quality assessment method of the above embodiment, and the system includes a preprocessing module, a pseudo-label generation module, a self-supervised sub-action parsing module, a motion sequence fine-grained feature expression extraction module, and a semi-supervised action quality assessment module;

[0086] Preprocessing module, used for video data preprocessing, including target detection and feature extraction in the video;

[0087] The pseudo-label generation module generates pseudo-labels for sub-actions in the teacher branch based on the teacher-student network;

[0088] The self-supervised sub-action parsing module is based on a teacher-student network and designs group contrastive learning to achieve self-supervised sub-action parsing of motion sequences;

[0089] The motion sequence fine-grained feature expression extraction module extracts the fine-grained feature expression of the motion sequence based on the self-supervised sub-action parsing module;

[0090] The semi-supervised action quality assessment module performs semi-supervised action quality assessment based on distillation learning.

[0091] It should also be noted that, in this specification, terms such as "comprises", "includes" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or apparatus comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or apparatus. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article or apparatus comprising the element.

[0092] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A semi-supervised action quality assessment method based on group contrastive learning, characterized by: The following steps are involved: S1, video data preprocessing, including target detection and feature extraction in the video; S2, based on the teacher-student network, generate sub-action pseudo labels in the teacher branch; S3. Based on the teacher-student network, we design group contrastive learning to achieve self-supervised sub-action parsing of motion sequences; S4, extracting fine-grained feature expressions of motion sequences based on self-supervised sub-action parsing; S5. Perform semi-supervised action quality evaluation based on distillation learning.

2. The semi-supervised action quality assessment method based on group contrastive learning according to claim 1 is characterized in that Step S1 is specifically as follows: Prepare a training set, including a set of labeled samples and the unlabeled sample set Among them, N l and N u Represents the number of labeled and unlabeled samples, y i Represents Corresponding real scoring data; X l and X u Used for subsequent semi-supervised action quality evaluation; Construct a teacher-student network, use a pre-trained target detector in the teacher branch, and l Detect human motion areas in video frames and extract the detected human motion areas to form a new sample set X u and X a This is then fed into a teacher-student network for training self-supervised sub-action parsing.

3. The semi-supervised action quality assessment method based on group contrastive learning according to claim 2 is characterized in that In the teacher-student network, the teacher branch uses the target detector to extract the motion area with concentrated information in the video through human target detection based on the initial video sample, and uses the I3D network to extract the set X a The teacher branch learns effective feature expressions that are not affected by scene changes based on the sample features, clusters the effective feature expressions, and gives sub-action pseudo labels based on the clustering results, which subsequently guide the student branch in sub-action analysis.

4. The semi-supervised action quality assessment method based on group contrastive learning according to claim 3 is characterized in that Generate pseudo labels for sub-actions in the teacher branch, specifically: Using the feature learning network, the attention mechanism is used to perform regional weighting on the video frame features to obtain representative feature expressions; in the teacher branch, the attention model is trained on the set X a The target area of the sample video frame after target detection is weighted, and the attention model after transfer training performs attention weighting on the video frame in the student branch to highlight the key motion information in the video frame; The attention weight calculation is specifically: The motion sequence features F extracted by the I3D network are w (t) Input the feature learning network, w∈{u,a}, perform global average pooling on each channel feature, and concatenate the pooled features to obtain a vector of dimension H×W×1; Perform convolution and Sigmoid activation operations on the pooled features to calculate the spatial attention weight, attention weight and video feature F of the video frame w (t) Perform multiplication operation to obtain the information-weighted features Where t represents the sequence time.

5. The semi-supervised action quality assessment method based on group contrastive learning according to claim 4 is characterized in that: Sub-action pseudo labels are generated by clustering the features, specifically: Features after weighted attention on the teacher branch Perform K-means clustering to obtain M cluster centers. Each cluster center corresponds to a pseudo label, and M sub-action pseudo labels are obtained. The process of selecting representative feature samples for M clusters adopts a greedy algorithm, specifically: Attention-weighted features And M cluster groups, select the significant samples representing each cluster from the M cluster groups, which are defined as the representative sample set representing the cluster center, and use S = {s m ,m=1,…,M} represents; The process of screening potential representative samples uses the following formula as the objective function and adopts a greedy algorithm to minimize the objective function: For the mth cluster center, select the first K samples that minimize the objective function Z(S) from the cluster sample group as the representative samples representing the mth cluster center, and store the K samples in the subset s m , is used as the representative sample of the mth sub-action, the set of M sub-action sample subsets is S = {s m ,m=1,…,M}; Encode M cluster centers, the mth sub-action set s m Assign pseudo label p m , the pseudo label set of M sub-actions is expressed as P = {p m ,m=1,…,M}.

6. The semi-supervised action quality assessment method based on group contrastive learning according to claim 5, characterized in that: Step S3 is specifically as follows: A sub-action classifier is constructed using a multi-layer perceptron and Softmax activation function. The sub-action sample subset S and pseudo-label set P selected from the teacher branch are used as training data for classifier training. The classifier is used to classify the sub-actions frame by frame in the video sequences of the teacher-student branches. The feature of each video frame is represented as The sub-action parsing process will feature Mapped into a probability vector, denoted as A={a m ,m=1,…,M}, represents the predicted value of the probability of the sub-action occurring in the video sequence, where a m The predicted probability corresponding to the mth sub-action; To ensure accurate prediction, self-supervised learning is used to keep the sub-action pseudo labels and sub-action parsing prediction values a m consistency between them; set a cross entropy loss function L c , using the trained classifier to classify features Perform sub-action prediction and obtain the predicted value Then use the sub-action pseudo label p m As a supervisory signal, by optimizing the cross entropy loss function L c , so that self-supervised learning maintains semantic consistency in feature learning and sub-action parsing; 7. The semi-supervised action quality assessment method based on group contrastive learning according to claim 6, characterized in that: Feature learning uses a multi-layer perceptron module to further refine feature expression, and uses a sub-action classifier to represent the feature Mapped to the sub-action category probability distribution vector h a (t),h u (t); The sub-action category probability distribution vector is further assigned a pseudo-label p corresponding to the category with the maximum activation value and high semantic similarity. Sub-actions with the same pseudo-label are grouped together. The process of grouping motion sequences into sub-actions is shown in the following formula: Among them, g(h w ) represents the calculation of h for a sub-action sequence w The average logical value of ,w∈{u,a}, represents the calculation of the two branches; It is an indicator function, which takes the value 1 when the condition is met, T B Indicates the number of sequences in a batch B; If two sub-action sequences in the teacher-student network branch correspond to the same sub-action pseudo-label at the same time, then the two sub-action sequences become a positive sample pair. Sub-action sequences from the same video group but classified into different pseudo-label groups are considered as negative sample pairs. By setting up group contrastive learning, the teacher-student network branches are made to bring the feature expressions of sub-actions of the same category closer and the feature expressions of different categories farther apart. The loss function of sub-action group contrastive learning is defined as: where D(·) denotes the distance calculation, τ is the temperature hyperparameter, and q,p∈P denotes different pseudo labels.

8. The semi-supervised action quality assessment method based on group contrastive learning according to claim 7, characterized in that: Step S4 is specifically as follows: Two network branches are set up to train supervised and unsupervised feature learning respectively; the supervised branch uses I3D network to extract action features and evaluate the action quality to obtain evaluation scores. These action features and evaluation score labels are used to train the supervised score regression model R for action quality evaluation. l (·), and then provide reference regression model parameters for unsupervised branch model training through distillation learning; the supervised regression loss function is defined as: Among them, y i Represents the true score label corresponding to sample i; In training the unsupervised branch of the score regression model R u (·), we first input the unlabeled data samples into the pre-trained regressor of the supervised branch to obtain the pseudo-labels of the unsupervised branch; by minimizing the difference between the prediction results of the unsupervised branch and the pseudo-labels generated by the supervised branch, we promote the unsupervised branch to learn features for accurate prediction of action scores and an accurate prediction score regressor; The semi-supervised motion quality assessment score regression loss function is conceptualized as follows: Among them, τ is the preset threshold; Represents the indicator function, which is 1 when the maximum predicted value exceeds τ, and 0 otherwise.

9. The semi-supervised action quality assessment method based on group contrastive learning according to claim 8, characterized in that: The overall training loss of the semi-supervised action quality assessment method is: L=L reg +λ1L sem +λ2L c +λ3L gc Among them, L reg is the supervised regression loss for supervised quality assessment, L sem Regularization loss for constrained pseudo-label consistency learning, L c is the self-supervised cross entropy loss, L gc is the sub-action group contrastive loss, and λ1, λ2, and λ3 are weight parameters.

10. A semi-supervised action quality assessment system based on group contrastive learning, characterized by: The system adopts the semi-supervised action quality assessment method according to any one of claims 1 to 9, and the system includes a preprocessing module, a pseudo-label generation module, a self-supervised sub-action parsing module, a motion sequence fine-grained feature expression extraction module, and a semi-supervised action quality assessment module; Preprocessing module, used for video data preprocessing, including target detection and feature extraction in the video; The pseudo-label generation module generates pseudo-labels for sub-actions in the teacher branch based on the teacher-student network; The self-supervised sub-action parsing module is based on a teacher-student network and designs group contrastive learning to achieve self-supervised sub-action parsing of motion sequences; The motion sequence fine-grained feature expression extraction module extracts the fine-grained feature expression of the motion sequence based on the self-supervised sub-action parsing module; The semi-supervised action quality assessment module performs semi-supervised action quality assessment based on distillation learning.

Images