Semi-supervised insulator image segmentation method and system based on binary teacher model

Abstract

The invention relates to the technical field of image processing, in particular to a semi-supervised insulator image segmentation method and system based on a binary teacher model. Comprising the steps of obtaining an insulator infrared image data set and dividing the insulator infrared image data set into a labeled image and an unlabeled image; constructing a semi-supervised segmentation model comprising a student model and a binary teacher model, wherein the student model comprises an encoder, a complementary Monte Carlo module and a feature enhancement decoder; calculating supervised loss of the labeled image; performing weak enhancement on the unlabeled image, then inputting the unlabeled image into a binary teacher model to generate a pseudo label, and performing fusion after class consistency and confidence check; performing strong enhancement on the label-free image, inputting the enhanced label-free image into a student model to generate complementary features, and calculating unsupervised loss; and updating parameters of the student model and the binary teacher model according to the total loss and then outputting a segmentation result. According to the method, the segmentation precision and the model generalization capability of the infrared image of the insulator are effectively improved under the condition of limited annotation data.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, specifically to a semi-supervised insulator image segmentation method and system based on a binary teacher model. Background Technology

[0002] Porcelain insulators, as one of the important pieces of equipment in power transmission lines, play a crucial role in electrical insulation and mechanical support. Due to the influence of environment and climate, porcelain insulators are prone to aging and failure, seriously threatening the safe operation of the power grid. Using infrared thermal imaging technology combined with image segmentation processing technology for fault diagnosis of porcelain insulators is currently a commonly used maintenance method.

[0003] Currently, most deep learning methods for insulator infrared image segmentation are based on fully supervised learning techniques. However, the effectiveness of fully supervised deep learning methods typically relies on large-scale, pixel-level labeled datasets. Acquiring high-quality labeled data is often costly and time-consuming. Due to limitations in acquisition conditions and annotation difficulty, the number of publicly available insulator infrared image segmentation datasets is very limited. Data acquisition and annotation require significant human and time resources, thus hindering the further application of fully supervised methods in this field. Furthermore, existing semi-supervised methods suffer from unstable pseudo-label generation and lack sufficient feature perception capabilities in insulator infrared images for feature extraction and fusion, affecting segmentation accuracy. Summary of the Invention

[0004] This invention provides a semi-supervised insulator image segmentation method and system based on a binary teacher model, aiming to solve the problem of existing fully supervised deep learning methods' dependence on large amounts of pixel-level labeled datasets, and improve the accuracy and generalization ability of insulator infrared image segmentation.

[0005] To achieve the above objectives, the present invention provides the following technical solution: This invention provides a semi-supervised insulator image segmentation method based on a binary teacher model, comprising: S1: Obtain the infrared image dataset of insulators and divide it into labeled images and unlabeled images; S2: Construct a semi-supervised infrared image segmentation model for insulators, which includes a student model and a binary teacher model. The student model includes an encoder, a complementary Monte Carlo module, and a feature enhancement decoder. The binary teacher model includes a first teacher model and a second teacher model. S3: Input the labeled image into the student model for prediction, and calculate the supervised loss based on the label; S4: After weak enhancement of the unlabeled image, input it into the first teacher model and the second teacher model respectively to obtain the first pseudo-label and the second pseudo-label. Perform class consistency check and confidence check on the first pseudo-label and the second pseudo-label. After retaining the pseudo-label that simultaneously satisfies class consistency and confidence is higher than the threshold, fuse them to obtain the fused pseudo-label. S5: After performing two strong enhancements on the unlabeled image to obtain two strongly enhanced images, they are input into the student model. Complementary features are generated through the complementary Monte Carlo module, and prediction results are generated through the feature enhancement decoder. The prediction results are supervised by the fused pseudo-label and the unsupervised loss is calculated. S6: Update the parameters of the student model and the binary teacher model according to the supervised loss and the unsupervised loss, and repeat steps S3 to S6 until training is complete; S7: Input the infrared image of the insulator to be segmented into the trained student model and output the segmentation result.

[0006] As a preferred technical solution of the present invention, in S3, after the labeled image has its feature map extracted by the encoder, it is input into the feature enhancement decoder to generate a prediction result.

[0007] As a preferred embodiment of the present invention, in step S4, the steps of category consistency check and confidence check include: The category prediction results of the first pseudo-label and the second pseudo-label at the same pixel position are compared, and only the pixel positions with the same category prediction results are retained; For the retained pixel locations, obtain the prediction confidence scores of the first teacher model and the second teacher model, and retain only the two pixel locations where both prediction confidence scores are higher than the threshold.

[0008] As a preferred embodiment of the present invention, in S4, the fusion step includes: For a pixel location that passes the category consistency check and the confidence check, the maximum value of the prediction confidence of the first teacher model and the second teacher model is used as the confidence of the fused pseudo-label at that pixel location.

[0009] In a preferred embodiment of the present invention, in S5, the two strongly enhanced images are processed by the student model to generate a first prediction result and a second prediction result, and the unsupervised loss is calculated based on the first prediction result, the second prediction result and the fused pseudo-label.

[0010] As a preferred embodiment of the present invention, in S5, the step of the complementary Monte Carlo module generating the complementary feature includes: Feature maps are extracted using the encoder; Generate a binary mask with the same shape and size as the feature map; The channels of the feature map are randomly divided into two complementary subsets; The binary mask is applied to the first subset to generate a first feature view; The complementary mask of the binary mask is applied to the second subset to generate a second feature view; The first feature view and the second feature view constitute the complementary feature.

[0011] As a preferred embodiment of the present invention, in S5, the feature enhancement decoder includes multiple cascaded fusion modules, and the processing steps of each fusion module include: The feature map from the upper decoder is concatenated with the feature map from the corresponding level of the encoder in the channel dimension; An attention weight map is generated from the concatenated feature map through a convolutional layer; The attention weight map is used to perform weighted fusion of the feature map from the upper decoder and the feature map from the corresponding level of the encoder; The fused feature maps are then subjected to residual convolution and upsampling processes in sequence before being output to the next layer fusion module.

[0012] As a preferred embodiment of the present invention, in S6, the parameters of the binary teacher model are updated using an exponential moving average method, and the parameters of the first teacher model and the second teacher model are updated according to the parameters of the student model.

[0013] As a preferred embodiment of the present invention, in S6, the parameter updates of the student model and the binary teacher model are based on the total loss, which is the weighted sum of the supervised loss and the unsupervised loss, and the weight of the unsupervised loss increases with the training period.

[0014] This invention also proposes a semi-supervised insulator image segmentation system based on a binary teacher model, comprising: The data partitioning module is used to acquire the infrared image dataset of insulators and divide it into labeled images and unlabeled images. The model building module is used to build a semi-supervised infrared image segmentation model for insulators, which includes a student model and a binary teacher model. The student model includes an encoder, a complementary Monte Carlo module, and a feature enhancement decoder. The binary teacher model includes a first teacher model and a second teacher model. The supervised training module is used to input labeled images into the student model for prediction and calculate supervised loss based on the labels; The pseudo-label generation module is used to perform weak enhancement on the unlabeled image and then input it into the first teacher model and the second teacher model to obtain the first pseudo-label and the second pseudo-label. The first pseudo-label and the second pseudo-label are subjected to category consistency check and confidence check. The pseudo-labels that simultaneously meet the category consistency and confidence level above the threshold are retained and then fused to obtain the fused pseudo-label. An unsupervised training module is used to perform two strong enhancements on the unlabeled image to obtain two strongly enhanced images, which are then input into the student model. Complementary features are generated through the complementary Monte Carlo module, and prediction results are generated through the feature enhancement decoder. The prediction results are supervised by the fused pseudo-label and the unsupervised loss is calculated. The parameter update module is used to update the parameters of the student model and the binary teacher model based on the supervised loss and the unsupervised loss. The segmentation inference module is used to input the infrared image of the insulator to be segmented into the trained student model and output the segmentation result.

[0015] The beneficial effects of this invention are: 1. This invention proposes a feature enhancement decoder, which adaptively weights and fuses feature maps from the upper-level decoder with feature maps from the corresponding level of the encoder through an attention weight mechanism. This can enhance the expression of key features and boundary information while suppressing irrelevant noise interference, and significantly improve the ability of semi-supervised models to perceive and extract features from infrared images of insulators.

[0016] 2. This invention proposes a binary teacher model and designs a pseudo-label consistency mechanism. By rigorously screening and fusing pseudo-labels generated by two independent teacher models through category consistency checks and confidence checks, the reliability of pseudo-labels is effectively improved. Combined with complementary features generated by the complementary Monte Carlo module, which emphasizes different textures and structures, the stability and robustness of model training are enhanced, significantly improving the segmentation performance of insulator infrared images under conditions of scarce labeled data. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart illustrating the semi-supervised insulator image segmentation method based on the binary teacher model of this invention. Figure 2 This is a structural diagram of a semi-supervised insulator infrared image segmentation model; Figure 3 It is a complementary Monte Carlo modular structure diagram; Figure 4 This is a diagram of the feature enhancement decoder model; Figure 5 This is a flowchart of the binary teacher pseudo-label screening process; Figure 6 This is a schematic diagram of the structure of the semi-supervised insulator image segmentation system based on the binary teacher model of the present invention. Detailed Implementation

[0018] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0019] Example 1: As Figure 1 As shown, the semi-supervised insulator image segmentation method based on the binary teacher model of the present invention includes: S1: Obtain the infrared image dataset of insulators and divide it into labeled images and unlabeled images; Specifically, infrared images of insulators are captured using an infrared camera, and some of the obtained infrared images of insulators are labeled to divide them into labeled data and unlabeled data, thus obtaining an infrared image dataset of insulators.

[0020] The image annotation process uses the Labelme annotation tool to outline the insulator contours in the infrared image as polygons, names it Insulator, saves it to generate a JSON label file, and then performs thresholding on the annotated image to obtain a more detailed label file.

[0021] S2: Construct a semi-supervised infrared image segmentation model for insulators, which includes a student model and a binary teacher model. The student model includes an encoder, a complementary Monte Carlo module, and a feature enhancement decoder. The binary teacher model includes a first teacher model and a second teacher model. Specifically, such as Figure 2 As shown, the semi-supervised insulator infrared image segmentation model mainly includes a student model and a binary teacher model.

[0022] The student model includes an encoder, a complementary Monte Carlo module, and a feature enhancement decoder. The encoder is used to extract multi-scale local features of the image. In a preferred embodiment, the encoder can be DINOV2_S, DINOV2_B, or DINOV2_L; for example... Figure 3 As shown, the complementary Monte Carlo module is used to create two feature maps from the feature map generated by the encoder, each emphasizing different textures and structures, for subsequent two-stream learning; as... Figure 4 As shown, the feature enhancement decoder is used to generate a prediction map corresponding to the feature map of the complementary Monte Carlo module.

[0023] The binary teacher model includes a first teacher model and a second teacher model. For example... Figure 5 As shown, the binary teacher model is used to generate pseudo-labels for weakly enhanced images of unlabeled images during semi-supervised training, which are then used to supervise the prediction maps generated by the student model and calculate the unsupervised loss.

[0024] S3: Input the labeled image into the student model for prediction, and calculate the supervised loss based on the label; Furthermore, in S3, after the labeled image has its feature map extracted by the encoder, it is input into the feature enhancement decoder to generate a prediction result.

[0025] Specifically, for labeled images, they are input into the student model for prediction. The labeled image first has its feature map extracted by the encoder, and then the extracted feature map is input into the feature enhancement decoder to generate a prediction result. The supervised loss for the labeled image is calculated based on manually provided labels.

[0026] The aforementioned supervised loss The calculation formula is as follows: ; in, This represents the number of labeled images. This is the index of the labeled image, with a value ranging from 1 to... ; For the first Student model prediction results with labeled images; For the first Zhang has the actual labels corresponding to the labeled images; Indicates the prediction result Softmax normalization is performed to obtain the probability distributions for each category.

[0027] S4: After weak enhancement of the unlabeled image, input it into the first teacher model and the second teacher model respectively to obtain the first pseudo-label and the second pseudo-label. Perform class consistency check and confidence check on the first pseudo-label and the second pseudo-label. After retaining the pseudo-label that simultaneously satisfies class consistency and confidence is higher than the threshold, fuse them to obtain the fused pseudo-label. Furthermore, the steps of the category consistency check and confidence check include: The category prediction results of the first pseudo-label and the second pseudo-label at the same pixel position are compared, and only the pixel positions with the same category prediction results are retained; For the retained pixel locations, obtain the prediction confidence scores of the first teacher model and the second teacher model, and retain only the two pixel locations where both prediction confidence scores are higher than the threshold.

[0028] Specifically, for unlabeled images, weak enhancement is first performed to generate weakly enhanced images. ,Will Input the first teacher model and the second teacher model respectively to generate the first pseudo-label. Second pseudo-label .

[0029] The class prediction results of the first pseudo-label and the second pseudo-label at the same pixel position are compared, and only pixel positions with the same class prediction result are retained. The judgment condition for class consistency check can be expressed as: ; in, Indicates pixel position; This indicates the pixel position of the first teacher model. The prediction category; Indicates the second teacher model at pixel position The prediction category; Weights for class consistency checks, applied when two teacher models predict the same class at that pixel location. The value is 1 if it is not 1, otherwise the value is 0.

[0030] For the retained pixel locations, the prediction confidence scores of the first teacher model and the second teacher model are obtained, and only pixel locations with two prediction confidence scores both higher than the threshold are retained. The determination criteria for the confidence score check can be expressed as: ; in, This indicates the pixel position of the first teacher model. The prediction confidence level; Indicates the second teacher model at pixel position The prediction confidence level; The preset confidence threshold for the binary teacher model; The weights for the confidence check are determined when the prediction confidence of both teacher models at that pixel location is higher than a threshold. hour, The value is 1 if it is not otherwise 0. In a preferred embodiment, the threshold is... Set it to 0.95.

[0031] Only when the pixel location passes both the class consistency check and the confidence check, i.e. and Only the pseudo-label corresponding to that pixel position is retained for subsequent training.

[0032] Furthermore, the fusion step includes: For a pixel location that passes the category consistency check and the confidence check, the maximum value of the prediction confidence of the first teacher model and the second teacher model is used as the confidence of the fused pseudo-label at that pixel location.

[0033] Specifically, the confidence score calculation formula for fused pseudo-labels is as follows: ; in, For the first teacher model at pixel position The prediction confidence level; For the second teacher model at pixel position The prediction confidence level; To merge pseudo-tags at pixel positions The final confidence level; This indicates the operation of retrieving the maximum value.

[0034] S5: After performing two strong enhancements on the unlabeled image to obtain two strongly enhanced images, they are input into the student model. Complementary features are generated through the complementary Monte Carlo module, and prediction results are generated through the feature enhancement decoder. The prediction results are supervised by the fused pseudo-label and the unsupervised loss is calculated. Furthermore, in S5, the two strongly enhanced images are processed by the student model to generate a first prediction result and a second prediction result, and the unsupervised loss is calculated based on the first prediction result, the second prediction result and the fused pseudo-label.

[0035] Specifically, for the weakly enhanced image in step S4 Perform two strong enhancement operations to obtain two strongly enhanced images. and .Will and The data are input into the student model and processed to generate the first prediction result. Second prediction results The unsupervised loss is calculated based on the first prediction result, the second prediction result, and the fused pseudo-label.

[0036] Further, in S5, the step of the complementary Monte Carlo module generating the complementary feature includes: Feature maps are extracted using the encoder; Generate a binary mask with the same shape and size as the feature map; The channels of the feature map are randomly divided into two complementary subsets; The binary mask is applied to the first subset to generate a first feature view; The complementary mask of the binary mask is applied to the second subset to generate a second feature view; The first feature view and the second feature view constitute the complementary feature.

[0037] Specifically, strongly enhanced image and Feature maps are generated separately using the encoder. and Generation and Feature Maps A binary mask M of the same shape and size will The channels are randomly divided into two complementary subsets. A binary mask M is applied to the first subset, and this operation is repeated twice to generate the first feature view; then a complementary mask is applied... The operation is applied to the second subset and repeated twice to generate a second feature view. The first and second feature views constitute complementary features. The above process can be represented as: ; ; in, and For weakly enhanced images Two strongly enhanced images obtained after performing two strong enhancement operations; and Strongly enhanced images and Feature maps generated by the encoder; For the generated and Binary masks of the same shape and size; To and Complementary masks; This represents an element-wise multiplication operation; This means repeating the operation twice.

[0038] Furthermore, in S5, the feature enhancement decoder includes multiple cascaded fusion modules, and the processing steps of each fusion module include: The feature map from the upper decoder is concatenated with the feature map from the corresponding level of the encoder in the channel dimension; An attention weight map is generated from the concatenated feature map through a convolutional layer; The attention weight map is used to perform weighted fusion of the feature map from the upper decoder and the feature map from the corresponding level of the encoder; The fused feature maps are then subjected to residual convolution and upsampling processes in sequence before being output to the next layer fusion module.

[0039] Specifically, the feature enhancement decoder mainly consists of four FE-Fusion modules. The feature map comes from the upper decoder path. This is a feature map processed by the complementary Monte Carlo module. First, [the following text is incomplete and likely refers to a separate process:] ... and The feature map is concatenated along the channel dimension to produce a feature map with double the number of channels. The concatenated feature map is then processed through a... The convolutional layer extracts the fusion weight information for each channel and each spatial location from the combined features, and then compresses the output value using the sigmoid activation function. Within the range, the attention weight map is obtained. The attention weight graph The calculation formula is as follows: ; in, The feature map comes from the upper decoder path; The feature map is derived from the complementary Monte Carlo module after processing. Indicates will and Perform splicing operations at the channel dimension; express Convolution operation; This indicates that the output value will be compressed to... Activation functions within the specified range; This is an attention weight map.

[0040] Obtain the attention weight map Then, using Feature maps from the upper-layer decoder and feature maps from complementary Monte Carlo modules Perform weighted fusion to obtain the fused feature map. The specific formula is as follows: ; in, This is an attention weight map; for Complementary weights; The feature map comes from the upper decoder path; The feature map is derived from the complementary Monte Carlo module after processing. The feature map obtained after weighted fusion; This indicates an element-wise multiplication operation.

[0041] The fused feature map F is processed sequentially through residual convolution units and upsampling, and then the feature dimension and number of channels are adjusted by projection operations. Finally, it is output to the next layer FE-Fusion module to generate the student model's prediction result for the image.

[0042] The unsupervised loss The calculation process is as follows: First Prediction Result Second prediction results Composed of complementary Monte Carlo modules and feature enhancement decoders The generation process is as follows: ; in, and Strongly enhanced images and Complementary features after processing by the complementary Monte Carlo module; Indicates a feature-enhanced decoder; for The corresponding first prediction result; for The corresponding second prediction result.

[0043] The formula for calculating unsupervised loss is as follows: ; ; in, This represents the number of unlabeled images. This is the index of the unlabeled image, with a value ranging from 1 to... ; For the first The first prediction result for the unlabeled image; For the first The second prediction result for the unlabeled image; These are the pseudo-tags that have been filtered through the pseudo-tag consistency mechanism; Represents the cross-entropy loss function; The weights for the unsupervised loss increase with the training period. This represents the current iteration cycle number; This represents the total number of iteration cycles.

[0044] S6: Update the parameters of the student model and the binary teacher model according to the supervised loss and the unsupervised loss, and repeat steps S3 to S6 until training is complete; Furthermore, in S6, the parameters of the binary teacher model are updated using an exponential moving average method, and the parameters of the first teacher model and the second teacher model are updated according to the parameters of the student model.

[0045] Furthermore, in S6, the parameter updates of the student model and the binary teacher model are based on the total loss, which is a weighted sum of the supervised loss and the unsupervised loss, and the weight of the unsupervised loss increases with the training period.

[0046] Specifically, the parameter updates of the student model and the binary teacher model are based on the total loss. The total loss is the weighted sum of the supervised loss and the unsupervised loss, calculated using the following formula: ; in, The supervised loss is calculated from the prediction results of the student model and the true labels in step S3. The unsupervised loss is calculated from the prediction results of the student model in step S5 and the fused pseudo-labels. The total loss is the sum of supervised and unsupervised losses. The weight of the unsupervised loss is... The calculation formula for the training cycle increases as the training cycle increases, as given in step S5.

[0047] Using the deep learning gradient backpropagation method, based on the total loss Update the parameters of the student model.

[0048] The parameters of the binary teacher model are updated using an exponential moving average method, updating the parameters of both the first and second teacher models based on the parameters of the student model. The update formula for the exponential moving average is as follows: ; ; in, Indicates the first Parameters of a teacher model ( As the first teacher model, (for the second teacher model) Represents the parameters of the student model; The momentum coefficient of the exponential moving average is used to control the fusion ratio of historical parameters and current student model parameters; The updated weights for the current student model parameters; This represents the current training cycle number.

[0049] Repeat steps S3 to S6 to continuously train and optimize until the preset number of training cycles is reached or the model converges, thus obtaining the trained student model.

[0050] S7: Input the infrared image of the insulator to be segmented into the trained student model and output the segmentation result.

[0051] Specifically, the infrared image of the insulator to be segmented is input into the trained student model. The image is then processed by an encoder to extract feature maps, and after being processed by a complementary Monte Carlo module, a feature enhancement decoder generates a prediction result, resulting in the segmented infrared image of the insulator.

[0052] In summary, this invention enhances the model's ability to perceive and extract insulator features through a feature enhancement decoder, optimizes pseudo-label quality through the pseudo-label consistency mechanism of the binary teacher model, and enhances the model's robustness by generating complementary features through a complementary Monte Carlo module. Under the condition of limited labeled data, it effectively improves the segmentation accuracy of insulator infrared images and the model's generalization ability.

[0053] Example 2: This invention also proposes a semi-supervised insulator image segmentation system based on a binary teacher model, such as... Figure 6 As shown, it includes a data partitioning module, a model building module, a supervised training module, a pseudo-label generation module, an unsupervised training module, a parameter update module, and a segmentation inference module.

[0054] To verify the effectiveness of the semi-supervised insulator image segmentation system based on the binary teacher model of this invention, the system was compared with several existing semi-supervised image segmentation methods. The experiment was based on a labeled insulator infrared dataset. After dividing the images into labeled and unlabeled segments using a data partitioning module, the system was trained sequentially through a model building module, a supervised training module, a pseudo-label generation module, an unsupervised training module, and a parameter update module. Finally, the segmentation inference module output the segmentation results. Compared with other comparative methods, the system of this invention achieved the best IoU results under all labeling ratio settings, especially at a labeling ratio of 1 / 16, indicating that the system of this invention can significantly improve the segmentation performance of insulator infrared images under conditions of scarce labeled data. The experimental results are shown in Table 1.

[0055] Table 1. Results of quantitative evaluation using different methods on the insulator infrared dataset.

[0056] As can be seen from Table 1, the system of the present invention outperforms existing semi-supervised image segmentation methods under different proportions of labeled images. In particular, when the labeled data is small (1 / 16 proportion), it improves by about 2.93 percentage points compared with the second-best method UniMatchV2, which fully demonstrates the effectiveness and superiority of the system of the present invention.

[0057] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A semi-supervised insulator image segmentation method based on a binary teacher model, characterized in that, include: S1: Obtain the infrared image dataset of insulators and divide it into labeled images and unlabeled images; S2: Construct a semi-supervised infrared image segmentation model for insulators, which includes a student model and a binary teacher model. The student model includes an encoder, a complementary Monte Carlo module, and a feature enhancement decoder. The binary teacher model includes a first teacher model and a second teacher model. S3: Input the labeled image into the student model for prediction, and calculate the supervised loss based on the label; S4: After weak enhancement of the unlabeled image, input it into the first teacher model and the second teacher model respectively to obtain the first pseudo-label and the second pseudo-label. Perform class consistency check and confidence check on the first pseudo-label and the second pseudo-label. After retaining the pseudo-label that simultaneously satisfies class consistency and confidence is higher than the threshold, fuse them to obtain the fused pseudo-label. S5: After performing two strong enhancements on the unlabeled image to obtain two strongly enhanced images, they are input into the student model. Complementary features are generated through the complementary Monte Carlo module, and prediction results are generated through the feature enhancement decoder. The prediction results are supervised by the fused pseudo-label and the unsupervised loss is calculated. S6: Update the parameters of the student model and the binary teacher model according to the supervised loss and the unsupervised loss, and repeat steps S3 to S6 until training is complete; S7: Input the infrared image of the insulator to be segmented into the trained student model and output the segmentation result.

2. The semi-supervised insulator image segmentation method based on a binary teacher model according to claim 1, characterized in that, In S3, after the labeled image has its feature map extracted by the encoder, it is input into the feature enhancement decoder to generate a prediction result.

3. The semi-supervised insulator image segmentation method based on a binary teacher model according to claim 1, characterized in that, In S4, the steps of the category consistency check and confidence check include: The category prediction results of the first pseudo-label and the second pseudo-label at the same pixel position are compared, and only the pixel positions with the same category prediction results are retained; For the retained pixel locations, obtain the prediction confidence scores of the first teacher model and the second teacher model, and retain only the two pixel locations where both prediction confidence scores are higher than the threshold.

4. The semi-supervised insulator image segmentation method based on a binary teacher model according to claim 3, characterized in that, In S4, the fusion step includes: For a pixel location that passes the category consistency check and the confidence check, the maximum value of the prediction confidence of the first teacher model and the second teacher model is used as the confidence of the fused pseudo-label at that pixel location.

5. The semi-supervised insulator image segmentation method based on a binary teacher model according to claim 1, characterized in that, In S5, the two strongly enhanced images are processed by the student model to generate a first prediction result and a second prediction result. The unsupervised loss is calculated based on the first prediction result, the second prediction result, and the fused pseudo-label.

6. The semi-supervised insulator image segmentation method based on a binary teacher model according to claim 5, characterized in that, In S5, the step of the complementary Monte Carlo module generating the complementary feature includes: Feature maps are extracted using the encoder; Generate a binary mask with the same shape and size as the feature map; The channels of the feature map are randomly divided into two complementary subsets; The binary mask is applied to the first subset to generate a first feature view; The complementary mask of the binary mask is applied to the second subset to generate a second feature view; The first feature view and the second feature view constitute the complementary feature.

7. The semi-supervised insulator image segmentation method based on a binary teacher model according to claim 1, characterized in that, In S5, the feature enhancement decoder includes multiple cascaded fusion modules, and the processing steps of each fusion module include: The feature map from the upper decoder is concatenated with the feature map from the corresponding level of the encoder in the channel dimension; An attention weight map is generated from the concatenated feature map through a convolutional layer; The attention weight map is used to perform weighted fusion of the feature map from the upper decoder and the feature map from the corresponding level of the encoder; The fused feature maps are then subjected to residual convolution and upsampling processes in sequence before being output to the next layer fusion module.

8. The semi-supervised insulator image segmentation method based on a binary teacher model according to claim 1, characterized in that, In S6, the parameters of the binary teacher model are updated using an exponential moving average method, and the parameters of the first teacher model and the second teacher model are updated according to the parameters of the student model.

9. The semi-supervised insulator image segmentation method based on a binary teacher model according to claim 1, characterized in that, In S6, the parameter updates of the student model and the binary teacher model are based on the total loss, which is a weighted sum of the supervised loss and the unsupervised loss, and the weight of the unsupervised loss increases with the training period.

10. A semi-supervised insulator image segmentation system based on a binary teacher model, characterized in that, The system is used to execute the semi-supervised insulator image segmentation method based on the binary teacher model as described in any one of claims 1-9, and the system comprises: The data partitioning module is used to acquire the infrared image dataset of insulators and divide it into labeled images and unlabeled images. The model building module is used to build a semi-supervised infrared image segmentation model for insulators, which includes a student model and a binary teacher model. The student model includes an encoder, a complementary Monte Carlo module, and a feature enhancement decoder. The binary teacher model includes a first teacher model and a second teacher model. The supervised training module is used to input labeled images into the student model for prediction and calculate supervised loss based on the labels; The pseudo-label generation module is used to perform weak enhancement on the unlabeled image and then input it into the first teacher model and the second teacher model to obtain the first pseudo-label and the second pseudo-label. The first pseudo-label and the second pseudo-label are subjected to category consistency check and confidence check. The pseudo-labels that simultaneously meet the category consistency and confidence level above the threshold are retained and then fused to obtain the fused pseudo-label. An unsupervised training module is used to perform two strong enhancements on the unlabeled image to obtain two strongly enhanced images, which are then input into the student model. Complementary features are generated through the complementary Monte Carlo module, and prediction results are generated through the feature enhancement decoder. The prediction results are supervised by the fused pseudo-label and the unsupervised loss is calculated. The parameter update module is used to update the parameters of the student model and the binary teacher model based on the supervised loss and the unsupervised loss. The segmentation inference module is used to input the infrared image of the insulator to be segmented into the trained student model and output the segmentation result.

Images