Semi-supervised incremental cigarette segmentation method based on large model semantic prior guidance
By adopting a semi-supervised method of semantic prior guidance in the class incremental smoke segmentation method, the combination of unlabeled data and labeled data is solved, and the existing methods are difficult to effectively utilize unlabeled data, realizing the continuous learning and retention of the characteristics of new and old smoke types by the model, improving the performance and robustness of the model.
Patent Information
- Application Number
- CN202411878308.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-19
- Publication Date
- 2025-04-04
AI Technical Summary
The existing incremental smoke segmentation method is difficult to effectively utilize unlabeled data in practical applications, which limits the learning ability and generalization performance of the model, resulting in the occurrence of false positives and missed reports.
The semi-supervised incremental smoke segmentation method based on semantic prior guidance of the big model is adopted to extract features from the old teacher model and student model through weak enhancement of unlabeled images and labeled images, and the semantic prior diagram is obtained using the SAM model to calculate the correlation consistency loss and regional prediction consistency loss under semantic prior guidance, and the model is updated with weighted sum.
On the basis of using a small number of labeled samples and a large number of unlabeled samples, the model is continuously learning the characteristics of new category smoke, while retaining the characteristics of old category smoke, improving the learning ability and generalization performance of the model, and reducing the false positive and false negative rate.
Smart Images
Figure CN119762781A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of fire prevention, and predicts fire by identifying smoke in the early stage of fire, and in particular to a semi-supervised incremental smoke segmentation method based on large model semantic prior guidance. Background Art
[0002] Smoke segmentation aims to segment different types of smoke from complex backgrounds, which helps to identify and respond to smoke in a timely manner and reduce fire risks. In actual application scenarios, there are many types of smoke, such as smoke on grasslands, smoke in forests, smoke from straw burning, smoke from factory warehouse fires, smoke from high-rise residential fires, smoke from shopping mall fires, smoke from factory emissions, smoke from chemical reactions, etc. In intelligent monitoring and automation systems, accurate distinction between different types of smoke can provide more accurate information for emergency response and reduce the false alarm rate of fires. However, as the environment changes, the characteristics of smoke are constantly changing. How to enable the model to have the ability to continuously learn new types of smoke is a key issue. Incremental smoke segmentation aims to continuously learn the features of new types of smoke to segment new types of smoke, while retaining the features of old types of smoke learned in previous tasks. Incremental smoke segmentation faces the following challenges: (1) Smoke is highly variable in shape, color, and concentration, and its characteristics may vary significantly under different ambient lighting and background conditions; (2) The stability-plasticity dilemma, i.e., the model must be both plastic in order to continuously learn new categories of smoke features and stable in order to not forget old categories of smoke features; and (3) Background semantic drift, i.e., at the current stage, the smoke in the background area may be of the old category that has been learned before, or it may be of the new category that will be learned in the future.
[0003] To address the above challenges, existing incremental smoke segmentation methods can be roughly divided into three categories: (1) Replay-based methods usually retain a portion of previous old category samples or generate additional old category samples for incremental training to prevent the forgetting of old category smoke features; (2) Dynamic structure-based methods usually use two parallel networks to store old category smoke features and learn new category smoke features respectively; (3) Knowledge distillation-based methods alleviate the catastrophic forgetting problem by transferring the knowledge of the old model to the new model.
[0004] However, existing incremental smoke segmentation methods usually assume that the incremental dataset is fully labeled. However, in practical application scenarios, this assumption is often not true. Since the pixel-level labeling process is time-consuming and laborious, it is difficult to obtain high-quality fully labeled datasets in many cases, especially in dynamic and complex environments. The dataset that can be used to train the model usually consists of a small amount of fully labeled data and a large amount of unlabeled data. Existing incremental smoke segmentation models cannot make full use of these unlabeled data during training, which limits their learning ability and generalization performance, and easily leads to false positives and false negatives. Summary of the invention
[0005] To solve the above problems, the present invention discloses a semi-supervised incremental smoke segmentation method based on large model semantic prior guidance, which is based on a small number of labeled samples and a large number of unlabeled samples, and achieves the goal of retaining the smoke features of old categories while continuously learning the smoke features of new categories.
[0006] In order to achieve the above object, the present invention provides the following technical solutions:
[0007] A semi-supervised incremental smoke segmentation method based on large model semantic prior guidance includes the following steps:
[0008] (1) Input the weakly enhanced unlabeled image into the old teacher model and the new teacher model, input the strongly enhanced unlabeled image into the student model, and input the labeled image into the old teacher model and the student model, extract the intermediate layer feature map and the deep layer feature map respectively, and then use the classifier to predict and obtain the prediction result;
[0009] (2) Input the weakly enhanced unlabeled images and labeled images into the SAM (Segment Anything Model) model to extract the semantic prior map;
[0010] (3) (For unlabeled images) Based on the deep feature maps extracted by the new teacher model and the student model and the semantic prior map, calculate the semantically guided relevance consistency loss;
[0011] (4) (For both unlabeled and labeled images) Calculate the region prediction consistency loss based on the prediction results obtained by the new teacher model and the student model and the semantic prior map;
[0012] (5) All losses are weighted summed, gradient backpropagation is performed to update the student model, and the new teacher model is updated by exponential moving average. The updated new teacher model will be used to perform semantic segmentation on the test samples.
[0013] Furthermore, the calculation process of the intermediate layer feature map, the deep layer feature map and the prediction result in step (1) includes the following steps:
[0014] (11) Take an unlabeled image x u ∈R 3×H×W After strong data enhancement Aug(·) and weak data enhancement aug(·) operations respectively, we get the strongly enhanced unlabeled image x us and weakly enhanced unlabeled image x uw ;
[0015] (12) Weakly enhance the unlabeled image x uw and the labeled image x l Input into the old teacher model f ot , extract the intermediate layer feature map and Weakly enhance the unlabeled image x uw Input to the new teacher model f nt In the above example, we extract deep feature maps. The unlabeled image x will be strongly enhanced us and the labeled image x l Input into the student model f s In the above example, we extract the intermediate layer feature map and F t l,m , and deep feature maps and F t l,d ;
[0016] (13) Old teacher model f ot The classifiers are respectively used for the weakly enhanced unlabeled images x uw and the labeled image x l Make predictions and get prediction results and New Teacher Model nt The classifier for the weakly enhanced unlabeled image x uw Make predictions and get prediction results Student Model s The classifiers are respectively used for the strongly enhanced unlabeled images x us and the labeled image x l Make predictions and get prediction results and
[0017] Furthermore, the method for calculating the semantic prior map in step (2) includes the following steps (weakly enhancing the unlabeled image x uw For example, there is a labeled image x l similar):
[0018] (21) Weakly enhance the unlabeled image x uwInput into the SAM (Segment Anything Model) model and set the SAM model to segmentation everything mode to obtain the SE diagram
[0019] (22) Based on SE graph You can get the mask of each target in the image (where K represents the number of targets), these target masks can be used as semantic prior maps. It should be noted that the target masks obtained here are category agnostic, that is, they only have spatial location information but no category information.
[0020] Furthermore, the calculation process of the semantically guided relevance consistency loss in step (3) includes the following steps:
[0021] (31) Based on the semantic prior map, the mask-guided semantic features can be obtained as follows:
[0022]
[0023] Among them, ⊙ represents the dot product operation, and P represents the average pooling operation;
[0024] (32) Based on the semantic features, the semantic correlation graph is calculated according to the following formula:
[0025]
[0026] Among them, ‖·‖2 represents the l2 norm;
[0027] (33) The semantically guided relevance consistency loss is calculated according to the following formula:
[0028]
[0029] Furthermore, the calculation process of the regional prediction consistency loss in step (4) includes the following steps (using the weakly enhanced unlabeled image x uw For example, there is a labeled image x l similar):
[0030] (41) Based on the student model f s For the strongly enhanced unlabeled image x us The prediction result is The pseudo label of the strongly enhanced unlabeled image is obtained according to the following formula:
[0031]
[0032] (42) Based on the target mask, the pseudo label can be Divided into different areas as follows:
[0033]
[0034] (43) Statistics about the categories are collected in each different region as follows:
[0035]
[0036] Where c∈C 0:t represents the cth category;
[0037] (44) Use Define a region R k The probability of belonging to category c (where |·| represents the number of pixels in the set). Ideally, the student model is for region R k The prediction results of all pixels in should be consistent, that is: is a one-hot vector. To this end, the regional prediction consistency loss is calculated according to the following formula:
[0038]
[0039] Furthermore, the calculation process of the overall loss in step (5) includes the following steps:
[0040] (51) According to the student model f s For a labeled image x l The prediction results and the true label Calculate the supervision loss L sup , as shown below:
[0041]
[0042] Among them, L ce represents the cross entropy loss;
[0043] (52) Based on and The category-specific threshold is updated by exponential moving average, and the incremental term Δτ is calculated according to the following formula: c :
[0044]
[0045] in, represents an indicator function (if the condition in the brackets is met, the function value is 1, otherwise, it is 0), [·] c Indicates obtaining the value of the cth channel;
[0046] (53) Based on the new teacher model fnt For the weakly enhanced unlabeled image x uw The prediction results Compute the pseudo labels for the weakly enhanced unlabeled images as follows:
[0047]
[0048] in, represents the prediction confidence, τ c represents the class-specific threshold;
[0049] (54) According to the student model f s For the strongly enhanced unlabeled image x us The prediction results and weakly enhanced pseudo labels for unlabeled images Calculate the consistency loss L cons , as shown below:
[0050]
[0051] (55) According to the old teacher model f ot For the weakly enhanced unlabeled image x uw The prediction results and student model f s For the strongly enhanced unlabeled image x us The prediction results Calculate the knowledge distillation loss of the output layer As shown below:
[0052]
[0053] Among them, L ukd Represents the unbiased knowledge distillation loss. Similarly, according to the old teacher model f ot For a labeled image x l The prediction results and student model f s For a labeled image x l The prediction results Calculate the knowledge distillation loss of the output layer As shown below:
[0054]
[0055] (56) According to the old teacher model f ot For the weakly enhanced unlabeled image x uw The extracted intermediate layer feature map and student model f s For the strongly enhanced unlabeled image x us The extracted intermediate layer feature map Calculate feature layer knowledge distillation loss As shown below:
[0056]
[0057] Among them, L mse Represents the MSE loss. Similarly, according to the old teacher model f ot For a labeled image x l The extracted intermediate layer feature map and student model f s For a labeled image x l The extracted intermediate layer feature map f t l,m Calculate feature layer knowledge distillation loss As shown below:
[0058]
[0059] (57) The weighted sum of each loss item is obtained to obtain the overall loss, as shown below:
[0060] L=L sup +L cons +λ kd L kd +λ sgcc L sgcc +λ rpc L rpc
[0061] Among them, λ kd , sgcc and λ rpc Represents the weight coefficient of each loss term, L kd =L lkd +L fkd ,
[0062] Compared with the prior art, the present invention has the following advantages and beneficial effects: (1) The present invention proposes a semi-supervised incremental smoke segmentation framework, which can use a small number of labeled samples and a large number of unlabeled samples to promote the model's learning of different categories of smoke features; (2) The present invention proposes a semantic-guided correlation consistency strategy to further explore the rich smoke feature representation in unlabeled samples; (3) The present invention proposes a regional prediction consistency strategy to make the model's predictions in a specific area tend to be consistent, thereby alleviating the problem of classification confusion. Combining the above advantages, the technical invention makes up for the defect that the existing method cannot effectively utilize unlabeled samples, and based on a small number of labeled samples and a large number of unlabeled samples, it achieves the purpose of retaining the smoke features of the old category while continuously learning the smoke features of the new category. BRIEF DESCRIPTION OF THE DRAWINGS
[0063] Figure 1 It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION
[0064] The technical solution provided by the present invention will be described in detail below in conjunction with specific embodiments. It should be understood that the following specific implementation methods are only used to illustrate the present invention and are not used to limit the scope of the present invention.
[0065] The present invention provides a semi-supervised incremental smoke segmentation method based on large model semantic prior guidance, and its flow chart is as follows: Figure 1 As shown, the following steps are included:
[0066] Step 1: Input the weakly enhanced unlabeled image into the old teacher model and the new teacher model, input the strongly enhanced unlabeled image into the student model, and input the labeled image into the old teacher model and the student model, extract the intermediate layer feature map and the deep layer feature map respectively, and then use the classifier to predict the prediction result.
[0067] The calculation process of the intermediate feature map, deep feature map and prediction results includes the following steps:
[0068] Step 1.1: Take an unlabeled image x u ∈R 3×H×W After strong data enhancement Aug(·) and weak data enhancement aug(·) operations respectively, we get the strongly enhanced unlabeled image x us and weakly enhanced unlabeled image x uw ;
[0069] Step 1.2: Weakly enhance the unlabeled image x uw and the labeled image x l Input into the old teacher model f ot , extract the intermediate layer feature map and Weakly enhance the unlabeled image x uw Input to the new teacher model f nt In the above example, we extract deep feature maps. The unlabeled image x will be strongly enhanced us and the labeled image x l Input into the student model f s In the above example, we extract the intermediate layer feature map and F t l,m , and deep feature maps and F t l,d ;
[0070] Step 1.3: Old teacher model f ot The classifiers are respectively used for the weakly enhanced unlabeled images xuw and the labeled image x l Make predictions and get prediction results and New Teacher Model nt The classifier for the weakly enhanced unlabeled image x uw Make predictions and get prediction results Student Model s The classifiers are respectively used for the strongly enhanced unlabeled images x us and the labeled image x l Make predictions and get prediction results and
[0071] Step 2: Input the weakly enhanced unlabeled image and the labeled image into the SAM (Segment Anything Model) model to extract the semantic prior map.
[0072] The calculation method of the semantic prior map includes the following steps (weakly enhancing the unlabeled image x uw For example, there is a labeled image x l similar):
[0073] Step 2.1: Weakly enhance the unlabeled image x uw Input into the SAM (Segment Anything Model) model and set the SAM model to segmentation everything mode to obtain the SE diagram
[0074] Step 2.2: Based on SE graph You can get the mask of each target in the image (where K represents the number of targets), these target masks can be used as semantic prior maps. It should be noted that the target masks obtained here are category agnostic, that is, they only have spatial location information but no category information.
[0075] Step 3: (For unlabeled images) Based on the deep feature maps extracted by the new teacher model and the student model and the semantic prior map, calculate the semantically guided relevance consistency loss.
[0076] The calculation process of semantically guided relevance consistency loss includes the following steps:
[0077] Step 3.1: Based on the semantic prior map, the mask-guided semantic features can be obtained as follows:
[0078]
[0079]
[0080] Among them, ⊙ represents the dot product operation, and P represents the average pooling operation;
[0081] Step 3.2: Based on the semantic features, calculate the semantic correlation graph according to the following formula:
[0082]
[0083] Among them, ‖·‖2 represents the l2 norm;
[0084] Step 3.3: Calculate the semantically guided relevance consistency loss according to the following formula:
[0085]
[0086] Step 4: (For both unlabeled and labeled images) Calculate the region prediction consistency loss based on the predictions obtained by the new teacher model and the student model and the semantic prior map.
[0087] The calculation process of regional prediction consistency loss includes the following steps (weakly enhancing the unlabeled image x uw For example, there is a labeled image x l similar):
[0088] Step 4.1: Based on the student model f s For the strongly enhanced unlabeled image x us The prediction result is The pseudo label of the strongly enhanced unlabeled image is obtained according to the following formula:
[0089]
[0090] Step 4.2: Based on the target mask, the pseudo label can be Divided into different areas as follows:
[0091]
[0092] Step 4.3: Collect statistics about the categories in each different region as follows:
[0093]
[0094] Where c∈C 0:t represents the cth category;
[0095] Step 4.4: Use Define a region R k The probability of belonging to category c (where |·| represents the number of pixels in the set). Ideally, the student model is for region Rk The prediction results of all pixels in should be consistent, that is: is a one-hot vector. To this end, the regional prediction consistency loss is calculated according to the following formula:
[0096]
[0097] Step 5: Perform weighted summation of all losses, perform gradient backpropagation to update the student model, and update the new teacher model by exponential moving average. The updated new teacher model will be used to perform semantic segmentation on the test sample.
[0098] The calculation process of the overall loss includes the following steps:
[0099] Step 5.1: According to the student model f s For a labeled image x l The prediction results and the true label Calculate the supervision loss L sup , as shown below:
[0100]
[0101] Among them, L ce represents the cross entropy loss;
[0102] Step 5.2: Based on and The category-specific threshold is updated by exponential moving average, and the incremental term Δτ is calculated according to the following formula: c :
[0103]
[0104] in, represents an indicator function (if the condition in the brackets is met, the function value is 1, otherwise, it is 0), [·] c Indicates obtaining the value of the cth channel;
[0105] Step 5.3: Based on the new teacher model f nt For the weakly enhanced unlabeled image x uw The prediction results Compute the pseudo labels for the weakly enhanced unlabeled images as follows:
[0106]
[0107] in, represents the prediction confidence, τ c represents the class-specific threshold;
[0108] Step 5.4: According to the student model f s For the strongly enhanced unlabeled image x us The prediction results and weakly enhanced pseudo labels for unlabeled images Calculate the consistency loss L cons , as shown below:
[0109]
[0110] Step 5.5: According to the old teacher model f ot For the weakly enhanced unlabeled image x uw The prediction results and student model f s For the strongly enhanced unlabeled image x us The prediction results Calculate the knowledge distillation loss of the output layer As shown below:
[0111]
[0112] Among them, L ukd Represents the unbiased knowledge distillation loss. Similarly, according to the old teacher model f ot For a labeled image x l The prediction results and student model f s For a labeled image x l The prediction results Calculate the knowledge distillation loss of the output layer As shown below:
[0113]
[0114] Step 5.6: According to the old teacher model f ot For the weakly enhanced unlabeled image x uw The extracted intermediate layer feature map and student model f s For the strongly enhanced unlabeled image x us The extracted intermediate layer feature map Calculate feature layer knowledge distillation loss As shown below:
[0115]
[0116] Among them, L mse Represents the MSE loss. Similarly, according to the old teacher model f ot For a labeled image x l The extracted intermediate layer feature map and student model f s For a labeled image xl The extracted intermediate layer feature map F t l,m Calculate feature layer knowledge distillation loss As shown below:
[0117]
[0118] Step 5.7: Perform a weighted sum of the various loss terms to obtain the overall loss, as shown below:
[0119] L=L sup +L cons +λ kd L kd +λ sgcc L sgcc +λ rpc L rpc
[0120] Among them, λ kd , sgcc and λ rpc Represents the weight coefficient of each loss term, L kd =L lkd +L fkd ,
[0121] The technical means disclosed in the scheme of the present invention are not limited to the technical means disclosed in the above-mentioned implementation mode, but also include technical schemes composed of any combination of the above-mentioned technical features. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications are also regarded as the protection scope of the present invention.
Claims
1. A semi-supervised incremental smoke segmentation method based on large model semantic prior guidance, characterized in that: The steps include: (1) Input the weakly enhanced unlabeled image into the old teacher model and the new teacher model, input the strongly enhanced unlabeled image into the student model, and input the labeled image into the old teacher model and the student model, extract the intermediate layer feature map and the deep layer feature map respectively, and then use the classifier to predict and obtain the prediction result; (2) Input the weakly enhanced unlabeled image and the labeled image into the SAM (Segment Anything Model) model to extract the semantic prior map; (3) (For unlabeled images) Based on the deep feature maps extracted by the new teacher model and the student model and the semantic prior map, calculate the semantically guided relevance consistency loss; (4) (For both unlabeled and labeled images) Calculate the region prediction consistency loss based on the prediction results obtained by the new teacher model and the student model and the semantic prior map; (5) All losses are weighted summed, gradient backpropagation is performed to update the student model, and the new teacher model is updated by exponential moving average. The updated new teacher model will be used to perform semantic segmentation on the test samples.
2. The semi-supervised incremental smoke segmentation method based on large model semantic prior guidance according to claim 1 is characterized in that: The calculation process of the intermediate layer feature map, the deep layer feature map and the prediction result in step (1) includes the following steps: (11) Take an unlabeled image x u ∈R 3×H×W After strong data enhancement Aug(·) and weak data enhancement aug(·) operations respectively, we get the strongly enhanced unlabeled image x us and weakly enhanced unlabeled image x uw ; (12) Weakly enhance the unlabeled image x uw and the labeled image x l Input into the old teacher model f ot , extract the intermediate layer feature map and Weakly enhance the unlabeled image x uw Input to the new teacher model f nt In the above example, we extract deep feature maps. The unlabeled image x will be strongly enhanced us and the labeled image x l Input into the student model f s In the above example, we extract the feature map of the intermediate layer. and And deep feature maps and (13) Old teacher model f ot The classifiers are respectively used for the weakly enhanced unlabeled images x uw and the labeled image x l Make predictions and get prediction results and New Teacher Model nt The classifier for the weakly enhanced unlabeled image x uw Make predictions and get prediction results Student Model s The classifiers are respectively used for the strongly enhanced unlabeled images x us and the labeled image x l Make predictions and get prediction results and 3. The semi-supervised incremental smoke segmentation method based on large model semantic prior guidance according to claim 1 is characterized in that: The method for calculating the semantic prior map in step (2) comprises the following steps (using the weakly enhanced unlabeled image x uw For example, there is a labeled image x l similar): (21) Weakly enhance the unlabeled image x uw Input into the SAM (Segment Anything Model) model and set the SAM model to segmentation everything mode to obtain the SE diagram (22) Based on SE graph You can get the (where K represents the number of targets), these target masks can be used as semantic prior maps. It should be noted that the target masks obtained here are category agnostic, that is, they only have spatial location information but no category information.
4. The semi-supervised incremental smoke segmentation method based on large model semantic prior guidance according to claim 1 is characterized in that: The calculation process of the semantically guided relevance consistency loss in step (3) includes the following steps: (31) Based on the semantic prior map, the mask-guided semantic features can be obtained as follows: Among them, ⊙ represents the dot product operation, and P represents the average pooling operation; (32) Based on the semantic features, the semantic correlation graph is calculated according to the following formula: Among them, ‖·‖2 represents the l2 norm; (33) The semantically guided relevance consistency loss is calculated according to the following formula:
5. The semi-supervised incremental smoke segmentation method based on large model semantic prior guidance according to claim 1 is characterized in that: The calculation process of the regional prediction consistency loss in step (4) includes the following steps (using the weakly enhanced unlabeled image x uw For example, there is a labeled image x l similar): (41) Based on the student model f s For the strongly enhanced unlabeled image x us The prediction result is The pseudo label of the strongly enhanced unlabeled image is obtained according to the following formula: (42) Based on the target mask, the pseudo label can be Divided into different areas as follows: (43) Statistics about the categories are collected in each different region as follows: Where c∈C 0:t represents the cth category; (44) Use Define a region R k The probability of belonging to category c (where |·| represents the number of pixels in the set). Ideally, the student model is for region R k The prediction results of all pixels in should be consistent, that is: is a one-hot vector. To this end, the regional prediction consistency loss is calculated according to the following formula:
6. The semi-supervised incremental smoke segmentation method based on large model semantic prior guidance according to claim 1 is characterized in that: The calculation process of the overall loss in step (5) includes the following steps: (51) According to the student model f s For a labeled image x l The prediction results and the true label Calculate the supervision loss L sup , as shown below: Among them, L ce represents the cross entropy loss; (52) Based on and The category-specific threshold is updated by exponential moving average, and the incremental term Δτ is calculated according to the following formula: c : in, represents an indicator function (if the condition in the brackets is met, the function value is 1, otherwise, it is 0), [·] c Indicates obtaining the value of the cth channel; (53) Based on the new teacher model f nt For the weakly enhanced unlabeled image x uw The prediction results Compute the pseudo labels for the weakly enhanced unlabeled images as follows: in, represents the prediction confidence, τ c represents the class-specific threshold; (54) According to the student model f s For the strongly enhanced unlabeled image x us The prediction results and weakly enhanced pseudo labels for unlabeled images Calculate the consistency loss L cons , as shown below: (55) According to the old teacher model f ot For the weakly enhanced unlabeled image x uw The prediction results and student model f s For the strongly enhanced unlabeled image x us The prediction results Calculate the knowledge distillation loss of the output layer As shown below: Among them, L ukd Represents the unbiased knowledge distillation loss. Similarly, according to the old teacher model f ot For a labeled image x l The prediction results and student model f s For a labeled image x l The prediction results Calculate the knowledge distillation loss of the output layer As shown below: (56) According to the old teacher model f ot For the weakly enhanced unlabeled image x uw The extracted intermediate layer feature map and student model f s For the strongly enhanced unlabeled image x us The extracted intermediate layer feature map Calculate feature layer knowledge distillation loss As shown below: Among them, L mse Represents the MSE loss. Similarly, according to the old teacher model f ot For a labeled image x l The extracted intermediate layer feature map and student model f s For a labeled image x l The extracted intermediate layer feature map Calculate feature layer knowledge distillation loss As shown below: (57) The weighted sum of each loss item is obtained to obtain the overall loss, as shown below: L=L sup +L cons +λ kd L kd +λ sgcc L sgcc +λ rpc L rpc Among them, λ kd , sgcc and λ rpc Represents the weight coefficient of each loss term, L kd =L lkd +L fkd ,