A method for detecting whether a judge's robe is worn in a court in a standard manner

Abstract

The application provides a detection method for judging whether the wearing of a judge's robe in a court is standard, and belongs to the technical field of image processing. The method comprises the following steps: collecting a pattern sample of the robe and a key point data set of the judge in different postures; constructing and training a YOLOv11-Pose human key point detection model and a FastR-CNN feature extraction and classification model; collecting image data in a fixed area of the court, detecting the judge and the key points thereof by using the trained YOLOv11-Pose human key point detection model; cutting the detected judge area, and inputting the trained FastR-CNN feature extraction and classification model to perform feature extraction and classification determination; and continuously monitoring the wearing state by using a DeepSORT algorithm. The method can realize real-time monitoring of the judge, and timely detect whether the wearing of the robe of the judge in the court is in line with the standard.

Description

Technical Field

[0001] This invention relates to a method for detecting whether judges wear their robes in court in a standardized manner, and belongs to the field of image processing technology. Background Technology

[0002] With the continuous improvement of the judicial system and the increasing public demand for judicial fairness, the image management of the courts has become particularly important. As representatives of the court, judges' attire directly affects the overall image of the court and the public's trust in the judicial system. To ensure the seriousness and authority of the courts, and to reflect the principles of judicial fairness and neutrality, detailed regulations are generally in place regarding judges' attire. These regulations typically include requiring judges to change into their robes or court attire before entering the courtroom, wear a standard tie, and display insignia such as judicial badges to highlight their professional identity and judicial authority.

[0003] However, in practice, due to human negligence, busy schedules, or other reasons, some judges may wear their robes inappropriately. These issues include, but are not limited to, wrinkled robes, missing or incorrectly positioned judicial badges, and ties that do not conform to regulations. These seemingly minor details can negatively impact the authority and credibility of the court. Especially in publicly broadcast settings such as online trials, judges' attire is more easily magnified and scrutinized, potentially leading to unnecessary controversy and questioning.

[0004] To effectively address this issue, ensure judges' courtroom attire conforms to regulations, and improve the intelligence level of court management, traditional oversight methods often rely on manual checks. This approach is not only inefficient but also susceptible to subjective influences, making it difficult to guarantee the comprehensiveness and accuracy of oversight. Therefore, developing a technological means to automatically and in real-time detect whether judges' robes are worn correctly is particularly important. Summary of the Invention

[0005] The purpose of this invention is to provide a method for detecting whether judges wear their robes in court in a standardized manner. This method can monitor judges in real time, detect whether judges wear their robes in court in a standardized manner, and ensure the stability of the detection results.

[0006] To achieve the above objectives, the present invention employs the following technical solution:

[0007] Firstly, a method for detecting whether judges' robes are worn correctly in court is provided, including the following steps:

[0008] Collect pattern samples of judges' robes and key point datasets of judges in different poses, as well as FastR-CNN dataset;

[0009] The YOLOv11-Pose human key point detection model and the FastR-CNN feature extraction and classification model are constructed and trained.

[0010] Image data in the fixed area of the court is collected, and the YOLOv11-Pose human key point detection model trained is used to detect the judge and the key points thereof.

[0011] The detected judge area is cut, the trained FastR-CNN feature extraction and classification model is input, and the shoulder pattern, the judge badge and the tie area are subjected to feature extraction and classification determination through a cascade classifier.

[0012] The DeepSORT algorithm is used to track the whole process of the judge from entering the court to sitting down, and the wearing state is continuously monitored.

[0013] Preferably, the YOLOv11-Pose human key point detection model is based on the YOLOV11 model, and a multi-scale convolution branch and a channel attention mechanism are added between the fourth C3k2 layer and the SPPF module of the backbone network.

[0014] Preferably, the multi-scale convolution branch and the channel attention mechanism are as follows:

[0015] Group convolution with different kernel sizes is used to extract multi-scale features in parallel;

[0016] An SEWeight module is applied to each branch output to generate channel attention weights: two fully connected layers are used to generate channel attention weights, the generated channel attention weights are applied to the feature map, and the weighted feature map is fused with multi-scale weights after Softmax; wherein the first fully connected layer is used for channel dimension reduction, and uses a ReLU activation function, and the second fully connected layer is used for restoring the number of channels, and uses a Sigmoid activation function.

[0017] Preferably, the FastR-CNN feature extraction and classification model comprises a feature extraction network, an RPN network, an ROI pooling, and a fully connected classification network, the feature extraction network uses a ResNet50 network, and the output feature maps of the last three layers of the ResNet50 network are fused.

[0018] Preferably, the output feature maps of the last three layers of the ResNet50 network are fused, and the fusion is as follows:

[0019] The output feature maps of the last three layers of the ResNet50 network are respectively a deep feature map of 14x14, a shallow feature map of 56x56, and a semantic information feature map of 28x28.

[0020] The shallow feature map is subjected to discrete wavelet transform.

[0021] The deep feature map is up-sampled and added to the semantic information feature map;

[0022] The fused feature map is up-sampled and added to the shallow feature map, and then input into the RPN network after being activated by a nonlinear activation function.

[0023] Preferably, the shallow feature map is subjected to discrete wavelet transform in the following manner:

[0024] The input shallow feature map is divided into four sub-bands by discrete wavelet transform, and the high-frequency components and the low-frequency components are arranged in the sub-bands;

[0025] The sub-band composed of the high-frequency components and the low-frequency components is added to the sub-band composed of the low-frequency components and the high-frequency components to generate a high-frequency weight;

[0026] The high-frequency weight is subjected to Hadamard multiplication operation with the sub-band composed of the low-frequency components and the low-frequency components to obtain an attention matrix;

[0027] The sub-band composed of the low-frequency components and the low-frequency components is element-wise added to the attention matrix to obtain a weighted feature.

[0028] Preferably, the DeepSORT tracking specifically comprises the following steps:

[0029] The key points and the classification determination result are input into the DeepSORT algorithm, and the motion trajectory is predicted by a Kalman filter;

[0030] If the wearing is detected to be non-standard, the result is continuously tracked and updated within a preset time interval;

[0031] A visual report is generated, and the judge wearing compliance state is output.

[0032] Preferably, the key points include shoulders, elbows, knees and ankles.

[0033] In a second aspect, a detection device for detecting whether the wearing of a judge's robe in a court is standard is provided, which comprises a processor and a memory storing program instructions, and the processor is configured to execute the detection method for detecting whether the wearing of a judge's robe in a court is standard when the program instructions are executed.

[0034] In a third aspect, a computer readable storage medium is provided, which stores a computer program, and the program is executed by a processor to implement the detection method for detecting whether the wearing of a judge's robe in a court is standard.

[0035] The present application has the following advantages:

[0036] The method can monitor the judges in real time, detect whether the wearing of a judge's robe in a court is standard in time, and ensure the uniformity and standardization of the court dress.

[0037] By adopting the YOLOv11-Pose human key point detection model and the FastR-CNN feature extraction and classification model, and combining the data enhancement technology, the detection accuracy of the model on the detail areas of the robe (such as the shoulder pattern, the chest law badge, the necktie area, etc.) is improved.

[0038] The feature fusion is performed on the feature maps output by the last three layers of the ResNet50 network, and the training result of small size signs is effectively improved through the add fusion method, thereby improving the detection capability of the model on small size signs on the robe (such as the law badge, the necktie, etc.).

[0039] In the detection process of the specific area of the robe wearing, a cascade classifier is used to first classify in a large range and then finely classify the detail area, so as to ensure the accuracy of the determination result.

[0040] The DeepSORT algorithm is used to track the whole process of the judge from entering the courtroom to sitting down, so as to ensure that the wearing state of each judge can be continuously monitored, and the behavior of wearing not in accordance with the standard can be found and corrected in time.

[0041] The method realizes the automatic detection of the robe wearing standard of the judge in the courtroom, reduces the burden of manual supervision, and improves the intelligent level of the courtroom management.

[0042] By adopting the data enhancement technology and focusing on learning the detail areas of the robe, the robustness of the model is improved, and the stability of the detection result under different environments and different light conditions is ensured. BRIEF DESCRIPTION OF DRAWINGS

[0043] The accompanying drawings are included to provide a further understanding of the present application, and constitute a part of the specification, which together with the embodiments of the present application, serve to explain the present application, and do not constitute a limitation of the present application.

[0044] Figure 1 It is a flowchart of the method of the present application.

[0045] Figure 2 It is a schematic diagram of the network structure of the YOLOv11-Pose human key point detection model.

[0046] Figure 3 It is a schematic diagram of the multi-feature fusion process of the YOLOv11-Pose human key point detection model.

[0047] Figure 4 It is a schematic diagram of the feature fusion module.

[0048] Figure 5 It is a schematic diagram of the feature fusion process. DETAILED DESCRIPTION

[0049] With reference to the accompanying drawings, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of the present application.

[0050] Embodiment 1

[0051] As shown in the figure, a detection method for detecting whether the wearing of a judge's robe in a court is standard, comprising the following steps: Figures 1-2

[0052] S1: Collecting a pattern sample of the robe, a key point data set of the judge in different postures and a FastR-CNN data set.

[0053] The data set includes robe pattern samples of different styles and judges in different postures under different environments (including different light and angles).

[0054] S2: Constructing and training a YOLOv11-Pose human key point detection model and a FastR-CNN feature extraction and classification model.

[0055] During the training process, data enhancement techniques are used to improve the robustness of the model, and the details of the robe (such as the shoulder pattern, the chest badge and the tie area) are learned in detail.

[0056] S3: Collecting image data in a fixed area of the court, and using the trained YOLOv11-Pose human key point detection model to detect the judge and his key points;

[0057] S4: Cutting the detected judge area, inputting the trained FastR-CNN feature extraction and classification model, and extracting and classifying the features of the shoulder pattern, the badge and the tie area through a cascaded classifier;

[0058] S5: Using the DeepSORT algorithm to track the whole process of the judge from entering the court to sitting down, and continuously monitoring the wearing state.

[0059] As a refinement of the above embodiment, the specific operation steps of step S1 for constructing the data set are as follows:

[0060] S101: Collecting samples of the judge wearing the robe under different environments (including different light and angles) through a high-definition camera installed in the court, to ensure that the samples cover different light and angles and different styles and patterns of the robe worn by the judge;

[0061] ​S102: Construct a dataset for training the YOLOv11-Pose human key point detection model, accurately label the judge's body key points including shoulders, elbows, knees, and ankles, and ensure the accuracy of the labeling;

[0062] S103: Construct a dataset for training the feature extraction and classification model of FastR-CNN, cut the detected judge area, save the two shoulder patterns, chest badge, and tie area details, and classify these areas through labeling.

[0063] S104: Extract and save the trained weight file to ensure that the model can be quickly loaded and real-time detected in the inference stage.

[0064] As a refinement of the above embodiment, the YOLOv11-Pose human key point detection model in step S2 is based on the YOLOV11 model, and a multi-scale convolution branch and a channel attention mechanism are added between the fourth C3k2 layer of the backbone network and the SPPF module, as shown in Figure 3 .

[0065] Specifically, the multi-scale convolution branch and the channel attention mechanism are as follows:

[0066] Parallelly extract multi-scale features using group convolution with different kernel sizes ((3x3, 5x5, 7x7));

[0067] Apply the SEWeight module to each branch output to generate channel attention weights: generate channel attention weights through two fully connected layers, apply the generated channel attention weights to the feature map, and fuse the multi-scale weights after weighting the feature map through Softmax; wherein the first fully connected layer is used for channel dimension reduction, and the ReLU activation function is used, and the second fully connected layer is used for restoring the channel number, and the Sigmoid activation function is used.

[0068] Fuse the multi-scale weights through Softmax, output after weighted fusion of multi-scale features, and obtain multi-scale feature maps in the channel dimension. The input channel dimension of each branch of the input feature map is C, and through this way, more rich position information of the input tensor can be obtained, and multiple branches are used to process in multiple scales. By using multi-scale convolution kernels in the pyramid structure, different scale spatial information is generated, and the channel dimension of the input tensor is effectively extracted to extract different scale spatial information on each channel dimension feature map.

[0069] Specifically, the training process of the YOLOv11-Pose human key point detection model is as follows:

[0070] The dataset of human key point detection is converted into a key point dataset suitable for yolo format, the human key points of the judge on the labeled picture are marked, and the expressed json file is converted into a key point data suitable for yolo format, and finally the txt content is generated as [00.484810.478960.700790.778860.313080.705972], wherein [0] is the category of the frame, for example, [0] is the head category, and [0.484810.478960.700790.77886] is the normalized center point horizontal and vertical coordinates, width, and height of the frame, and [0.313080.705972] represents the normalized horizontal and vertical coordinates of the first key point, and the key point visibility, wherein the key point visibility [0] represents invisibility, [1] represents occlusion, and [2] represents visibility.

[0071] The YOLOv11-Pose model is trained using the constructed human key point dataset, and the key points of the judge, such as the shoulders, elbows, and knees, are focused on learning. This step needs to adjust the learning rate, batch size, and other hyperparameters according to the size and complexity of the training set to ensure that the model can stably converge.

[0072] As a refinement of the above embodiment, the FastR-CNN feature extraction and classification model in step S2 includes a feature extraction network, an RPN network, an ROI pooling, and a fully connected classification network. The feature extraction network uses a ResNet50 network. The last layer feature map output by the ResNet50 network is input to the RPN network to generate candidate regions, then the ROI pooling layer is used to adjust the candidate regions of different sizes, and finally the fully connected classification network is used to realize classification and region frame regression. In the above description of the model, only the deepest features in the feature extraction network are used. The deepest features have a large receptive field and are abstract, and after a series of convolution and pooling operations, many small size landmark detail information is lost, which makes the model not suitable for detecting small size landmarks, and thus inaccurate detection occurs. On the contrary, shallow features have complete small size landmark information due to fewer convolution and pooling operations. To effectively extract small size landmark features, deep and shallow features can be fused to improve the performance of the model.

[0073] Specifically, the output feature maps of the last three layers of the ResNet50 network are fused. The upsampling method is used to enlarge the feature map to twice the original size, and the add fusion method with small calculation amount is used for feature fusion. In the original FasterR-CNN model, the feature extraction network finally outputs a feature map with a size of 14x14. By fusing the feature maps of the last three layers, a feature map with a size of 56x56 is obtained. The fused feature map retains all the information of the feature maps output by the last three layers of the network, effectively improving the detection effect of small targets.

[0074] Specifically, the output feature maps of the last three layers of the ResNet50 network are fused. Since the spatial resolution of the shallow feature map is higher and the resolution of the deep feature map is lower, the deep feature map needs to be up-sampled by using the bilinear interpolation method to adjust the size of the deep feature map to the size of the shallow feature map, from 14x14 to 56x56.

[0075] As shown in Figure 5 , the method comprises the following steps:

[0076] (1) The output feature maps of the last three layers of the ResNet50 network are respectively a deep feature map of 14x14, a shallow feature map of 56x56, and a semantic information feature map of 28x28.

[0077] (2) Discrete wavelet transform is performed on the shallow feature map.

[0078] (3) The deep feature map is up-sampled to 28x28, and then added to the semantic information feature map.

[0079] (4) The fused feature map is up-sampled to 56x56, and then added to the shallow feature map. Then, the result is activated by a nonlinear activation function and input into an RPN network, so as to enhance the expression ability of the model to details.

[0080] The fused feature map output by the ResNet50 network is input into an RPN (Region Proposal Network) to generate candidate regions, and then the feature is integrated by an ROI pooling layer for final target classification and regression. After the RPN and the ROI pooling, the candidate regions are input into a fully connected classification network for target classification and regression.

[0081] Specifically, as shown in Figure 4 and Figure 5 , the discrete wavelet transform (DWT) is performed on the shallow feature map as follows:

[0082] The input shallow feature map is divided into four sub-bands by using the discrete wavelet transform, which are arranged by high-frequency components and low-frequency components.

[0083] ,

[0084] ,

[0085] Among them, represents the low-frequency component, represents the high-frequency component, represents the two-dimensional discrete wavelet transform, This represents a scaling function used to extract the low-frequency components of an image. Indicates the pixel position in which row of the image. This indicates the pixel position in which column of the image. This represents a wavelet function used to extract high-frequency information from an image. Indicates the position of a pixel. .

[0086] The high-frequency weight is generated by adding the subband composed of high-frequency and low-frequency components to the subband composed of low-frequency and high-frequency components.

[0087] ,

[0088] in, Indicates high-frequency weights, This represents the Sigmoid function. This indicates the number of channels, specifically the RGB color channels. Indicates altitude, Indicates width, Represents the set of real numbers

[0089] The attention matrix is ​​obtained by performing Hadamard multiplication on the high-frequency weights, low-frequency components, and subbands composed of low-frequency components. :

[0090] ,

[0091] Subband composed of low-frequency components The weighted features are obtained by adding the elements of the attention matrix together. :

[0092] .

[0093] DWT decomposes the input signal into four discrete wavelet sub-bands without loss of information and embeds detailed information into the output. Therefore, DWT ensures the integrity of the input features after pooling. It ensures that small targets are fully preserved and restored, while also ensuring that they are not affected by jagged edges and noise.

[0094] Specifically, the FastR-CNN feature extraction and classification model training process is as follows:

[0095] Using the pre-built FastR-CNN dataset, the model is trained to extract detailed features of the robes. These include: feature extraction of the shoulder patterns, chest insignia, and tie area of ​​the robes; the FastR-CNN model then classifies these detailed areas to determine whether they conform to the specifications.

[0096] During the training process, hyperparameters need to be adjusted through cross-validation methods to ensure the accuracy and robustness of the model.

[0097] As a refinement of the above embodiment, the step S3 specifically comprises the following steps:

[0098] S301: Crop the image of the judge's location, ensure that the cropped area accurately contains the judge and his surroundings, and avoid redundant data affecting subsequent processing;

[0099] S302: Send the cropped image to the YOLOv11-Pose human key point detection model to detect the key points of the judge. The model detects the judge in the image and identifies the key points, including the shoulders, elbows, knees, and ankles, to ensure that the model can accurately identify the judge's posture and movement;

[0100] S303: Save the detected judge key point information, including the coordinate position of each key point, and provide it to the FastR-CNN feature extraction and classification model;

[0101] S304: Visualize the judge tracking results and save the detection data to an output file for further analysis and review.

[0102] As a refinement of the above embodiment, the step S4 specifically comprises the following steps:

[0103] S401: Candidate region extraction, generate region candidates for the judge image to determine the regions that may contain robe features;

[0104] S402: Image preprocessing, resize each candidate region to ensure consistent size for input to FastR-CNN; pixel value normalization is also required to adapt to the input requirements of the model;

[0105] S403: Feature extraction, send the preprocessed candidate region to the FastR-CNN model to extract deep features through convolution layers. The model classifies and regresses each candidate region through fully connected layers to determine whether the robe region meets the specifications;

[0106] S404: Output classification results, determine whether the robe's shoulder pattern, chest badge, and necktie region meet the specifications based on the extracted features and classification results. If they meet the specifications, return the qualified result; otherwise, return the unqualified result;

[0107] S405: Store the detection results, store the detection results of each robe region as a data file for subsequent judge tracking and wear detection judgment.

[0108] As a refinement of the above embodiment, the specific steps of DeepSORT tracking in step S5 include:

[0109] S501: input the judge detection result, input the obtained judge key point and robe detection result into the DeepSORT algorithm for multi-target tracking; DeepSORT predicts the motion trajectory of the judge through the Kalman filter, and performs data association;

[0110] S502: judge motion trajectory tracking, in the courtroom environment, the judge may move or stand, DeepSORT accurately tracks the motion trajectory of each judge, identifies the identity of the judge in each frame of image, and keeps tracking until the judge sits down;

[0111] S503: wear standard judgment, according to the identity of the judge tracked by DeepSORT, combined with the classification result, judge whether the robe of the judge is in conformity with the standard, if the judge wears not standard, continuously monitor in a certain time, and update the result when the wear is restored to standard.

[0112] S504: output the detection result, continuously monitor the wear result of each judge, and update according to the time interval, generate a visual report, output the wear compliance of the judge and any non-standard behavior detected.

[0113] The following compares the data of multiple prior art and the technical solution in terms of basic performance index, accuracy and ablation experiment, as shown in Tables 1-3, it can be obtained that the technical solution has good effect in basic performance index, accuracy and ablation experiment, and the comprehensive performance is optimal.

[0114] Table 1: basic performance index

[0115]

[0116] Table 2: accuracy

[0117]

[0118] Table 3: ablation experiment

[0119]

[0120] Example 2

[0121] The embodiment of the present disclosure further provides a detection device for detecting whether the wearing of a judge's robe in a court is standard, comprising a processor and a memory. Optionally, the device can further comprise a communication interface and a bus. The processor, the communication interface and the memory can communicate with each other through the bus. The communication interface can be used for information transmission. The processor can call the logical instructions in the memory to execute the detection method for detecting whether the wearing of a judge's robe in a court is standard.

[0122] In addition, the logical instructions in the memory described above can be realized in the form of a software function unit and sold or used as an independent product, and can be stored in a computer readable storage medium.

[0123] The memory, as a computer readable storage medium, can be used to store software programs, computer executable programs, such as program instructions / modules corresponding to the method in the embodiment of the present disclosure. The processor executes the program instructions / modules stored in the memory, thereby executing the function application and data processing, that is, realizing the detection method for detecting whether the wearing of a judge's robe in a court is standard in the embodiment described above.

[0124] The memory can include a program storage area and a data storage area, wherein the program storage area can store an operating system and application programs required by at least one function; the data storage area can store data created according to the use of the terminal device, etc. In addition, the memory can include a high-speed random access memory and can also include a non-volatile memory.

[0125] The embodiment of the present disclosure provides a computer readable storage medium, which stores computer executable instructions, and the computer executable instructions are set to execute the detection method for detecting whether the wearing of a judge's robe in a court is standard.

[0126] The computer readable storage medium described above can be a transitory computer readable storage medium or a non-transitory computer readable storage medium.

[0127] The technical solution of the embodiment of the present disclosure can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes one or more instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the method described in the embodiment of the present disclosure. The aforementioned storage medium can be a non-transitory storage medium, including a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc. various media that can store program codes, or can be a transitory storage medium.

[0128] Finally, it should be noted that the above only describes the preferred embodiments of the present application and is not intended to limit the present application. Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art will appreciate that modifications can be made to the technical solutions described in the foregoing embodiments, or some of the technical features thereof can be replaced by equivalent features. Any modifications, equivalent replacements, improvements, and the like made within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. A method for detecting whether judges' robes are worn correctly in court, characterized in that, Includes the following steps: Collect pattern samples of judges' robes and key point datasets of judges in different poses, as well as FastR-CNN dataset; Construct and train the YOLOv11-Pose human keypoint detection model and the FastR-CNN feature extraction and classification model; Image data of a fixed area in the courtroom was collected, and the YOLOv11-Pose human keypoint detection model, which had been trained, was used to detect the judge and his key points. The detected judge area is cut out and input into the trained FastR-CNN feature extraction and classification model. The cascaded classifier is used to extract features and classify the shoulder pattern, judicial emblem, and tie area. The DeepSORT algorithm is used to track the judge's entire process from entering the courtroom to taking his seat, continuously monitoring his attire status; The YOLOv11-Pose human keypoint detection model is based on the YOLOv11 model, and adds a multi-scale convolutional branch and channel attention mechanism between the fourth C3k2 layer of the backbone network and the SPPF module. The FastR-CNN feature extraction and classification model includes a feature extraction network, an RPN network, ROI pooling, and a fully connected classification network. The feature extraction network uses a ResNet50 network, and the feature maps output by the last three layers of the ResNet50 network are fused. The feature maps output from the last three layers of the ResNet50 network are fused as follows: The output feature maps of the last three layers of the ResNet50 network are: a 14×14 deep feature map, a 56×56 shallow feature map, and a 28×28 semantic information feature map. Perform discrete wavelet transform on the shallow feature map; The deep feature map is upsampled and then fused with the semantic information feature map by addition. The fused feature map is upsampled and then added to the shallow feature map for fusion. Finally, it is activated by a nonlinear activation function and input into the RPN network.

2. The method for detecting whether judges' robes are worn correctly in court, as described in claim 1, is characterized in that... The multi-scale convolutional branching and channel attention mechanism are specifically as follows: Multi-scale features are extracted in parallel using group convolutions with different kernel sizes; For each branch output, the SEWeight module is applied to generate channel attention weights: channel attention weights are generated through two fully connected layers, and the generated channel attention weights are applied to the feature map. The feature map is then weighted and fused with multi-scale weights via Softmax for output. The first fully connected layer is used for channel dimensionality reduction using the ReLU activation function, and the second fully connected layer is used to restore the number of channels using the Sigmoid activation function.

3. The method for detecting whether judges' robes are worn correctly in court, as described in claim 1, is characterized in that... The specific method for performing discrete wavelet transform on shallow feature maps is as follows: The input shallow feature map is divided into four sub-bands using discrete wavelet transform, which are composed of high-frequency and low-frequency components respectively. The high-frequency weight is generated by adding the subband composed of high-frequency and low-frequency components to the subband composed of low-frequency and high-frequency components. The attention matrix is ​​obtained by performing Hadamard multiplication on the high-frequency weights, low-frequency components, and subbands composed of low-frequency components. The weighted features are obtained by adding the elements of the low-frequency components, the sub-bands composed of the low-frequency components, and the attention matrix.

4. The method for detecting whether judges' robes are worn correctly in court, as described in claim 1, is characterized in that... The specific steps of DeepSORT tracing include: The key points and classification results are input into the DeepSORT algorithm, and the motion trajectory is predicted by the Kalman filter. If improper wearing is detected, the system will continuously track the device and update the results at preset time intervals. Generate a visual report that displays the judge's attire compliance status.

5. The method for detecting whether judges' robes are worn correctly in court, as described in any one of claims 1-4, is characterized in that... The key points include the shoulders, elbows, knees, and ankles.

6. A detection device for standardizing the wearing of judges' robes in court, comprising a processor and a memory storing program instructions, characterized in that, The processor is configured to, when running the program instructions, execute the detection method for determining whether judges wear their robes properly in court as described in any one of claims 1-5.

7. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements a method for detecting whether a judge's robes are worn correctly in a courtroom, as described in any one of claims 1-5 above.

Images