Unsupervised domain incremental learning method guided by global channel prior

Abstract

The invention relates to the technical field related to automatic driving, in particular to an unsupervised domain incremental learning method guided by global channel prior, which comprises the following steps of: 1, extracting shallow layer features of a test image through a shallow layer pre-training layer of a baseline model, and 2, carrying out global average pooling on the shallow layer features to obtain channel aggregation features. According to the unsupervised domain incremental learning method guided by the global channel prior, through deep analysis of shallow representation distribution, channel aggregation features of the same domain tend to be clustered, cluster center distances between different domains are significant and irrelevant to image contents, the cluster center distances are defined as inter-domain channel prior, and meanwhile, the cluster center distances are defined as inter-domain channel prior. Channel aggregation features of image blocks with similar structures in the same domain present a consistent channel response mode, and different domains have significant differences, and intra-domain channel prior characterization is used.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of automatic driving, and in particular to an unsupervised domain incremental learning method guided by global channel prior. BACKGROUND

[0002] Robustness of intelligent perception model is crucial for intelligent systems such as autonomous driving and road video monitoring. However, the actual image acquisition process is often affected by factors such as bad weather and poor lighting, resulting in degraded images. In the representation space, there is a significant shift between degraded images and clear images, which is known as the "domain shift" phenomenon, and is one of the key factors affecting model robustness. Since traditional perception models only learn feature representations specific to training data during the training process, they are often not robust enough when facing continuous domain shift problems. However, as the real-time space evolves, models may encounter challenges of continuous domain changes, and the test data received by the model is usually unlabeled. Therefore, intelligent perception models must have sufficient robustness and do not rely on large amounts of labeled data to ensure stable performance.

[0003] To improve the adaptability of perception models, existing solutions are mainly divided into three categories: image restoration-based methods, unsupervised domain adaptation, and incremental learning. Image restoration-based methods usually focus on improving the visual effects of images, but cannot guarantee stable performance improvement. Unsupervised domain adaptation learns domain-invariant features to achieve model transfer from labeled source domain to unlabeled target domain, but this method cannot cope with continuous domain changes, and it is difficult to generalize to new domains and may also lose knowledge learned on the source domain. Incremental learning (IL) is an effective strategy to cope with catastrophic forgetting, but current incremental learning methods rely on access to task IDs during inference, which limits the flexibility of the model in real-world scenarios. SUMMARY

[0004] The present application aims to provide an unsupervised domain incremental learning method guided by global channel prior to solve the problem of continuous domain change challenge that cannot be effectively alleviated in the background technology.

[0005] To achieve the above purpose, the present application provides the following technical solution: an unsupervised domain incremental learning method guided by global channel prior, comprising the following steps:

[0006] Step one, extracting shallow features of the test image through the shallow pre-training layer of the baseline model;

[0007] Step two, performing global average pooling on the shallow features to obtain channel aggregated features;

[0008] Step three, based on the channel aggregation feature, the domain discriminant mechanism is used to judge the domain attribute of the test image, if the test image comes from the known domain, the corresponding dynamic channel response optimization module (DCROM) is loaded from the memory bank and integrated into the baseline model for testing, if the test image comes from the unknown domain, the training mode is switched to, the unpaired clear and test images are used for unsupervised adversarial training, and after the training is completed, the generated DCROM is stored in the memory bank;

[0009] Step four, in the training process, based on the global channel priori, the dynamic channel response optimization module is guided to learn the channel response mapping relationship between the degraded features and the clear features, and the channel response of the degraded features is dynamically optimized.

[0010] Preferably, the global channel priori includes:

[0011] Inter-domain channel priori: the channel aggregation features of the same domain tend to cluster, and the cluster center distance between different domains is significant, which is independent of image content;

[0012] Intra-domain channel priori: the channel aggregation features of the structure similar image blocks in the same domain present consistent channel response mode, while there is significant difference between different domains.

[0013] Preferably, the domain discriminant mechanism (DDM) includes the following steps:

[0014] 1) Extract the channel aggregation feature matrix of the test image;

[0015] 2) Dimension reduction processing is performed on the channel aggregation feature matrix;

[0016] 3) The center point coordinates of the dimension-reduced feature matrix are calculated, and the distance is compared with the feature center of the known domain;

[0017] 4) According to the minimum distance, it is judged whether the test image belongs to the known domain, if the minimum distance is greater than the preset threshold, it is determined as unknown domain, if the minimum distance is less than or equal to the preset threshold, it is determined as known domain, and the corresponding domain ID is assigned.

[0018] Preferably, the dynamic channel response optimization module (DCROM) includes:

[0019] Stage division and initial convolution processing: at the beginning of each stage, the input features pass through a layer of 3x3 convolution, which maps the channel dimension to a preset value;

[0020] Star Block operation: in each stage, the features are sequentially processed by multiple Star Blocks, wherein 1x1 convolution is used to promote linear combination of channel responses, star operation is introduced to fuse features of different subspaces, channel dimension is reduced to reduce redundancy and then increased to enhance nonlinear expression capability, and residual link is adopted to reduce feature loss.

[0021] Preferably, the unsupervised domain incremental learning method (UDIL) comprises:

[0022] The memory stores DCROM specific to each domain;

[0023] In the training stage, the DCROM is optimized by unsupervised adversarial training for unknown domain data, and is stored in the memory;

[0024] In the test stage, the corresponding DCROM is loaded according to the domain discrimination result, and is integrated into the baseline model for inference.

[0025] Preferably, the baseline model comprises any one of VGG16, ResNet50 or CMT.

[0026] Preferably, the method is verified by experiments on the synthetic degraded data set lmageNet-C, the real fog data set Haze20 and the real low-light data set ExDark, and the method is superior to the baseline model, the unsupervised image restoration method, the unified image restoration method and the test-time adaptive method in terms of Top-1 classification accuracy.

[0027] Preferably, the dynamic channel response optimization module (DCROM) only updates the weights of the generator and the discriminator in the training process, and the weights of the baseline model remain unchanged.

[0028] Preferably, the domain discrimination mechanism (DDM) can flexibly adapt to an unknown number and any type of domain without providing any prior information.

[0029] Preferably, the global channel prior is verified by statistics experiments, including channel aggregation feature distribution analysis of degraded images in clear domains, fog domains and snowy domains, and channel response analysis of image blocks under fog and low-light conditions.

[0030] Compared with the prior art, the present application has the following advantages:

[0031] 1. A global channel prior is proposed. Through in-depth analysis of the distribution of shallow representations, channel aggregation features in the same domain tend to cluster, and the distance between cluster centers between different domains is significant, regardless of the image content. This is defined as the inter-domain channel prior. At the same time, the channel aggregation features of image blocks with similar structures in the same domain show consistent channel response patterns, while there are significant differences between different domains, which are represented using the intra-domain channel prior.

[0032] 2. A domain discrimination mechanism (DDM) is proposed to quickly discriminate the domain attributes of test data. The mechanism can flexibly adapt to an unknown number and arbitrary types of domains without providing any prior information, providing an innovative solution for domain incremental learning.

[0033] 3. A lightweight dynamic channel response optimization module (DCROM) is proposed, which aims to learn the channel response mapping relationship between degraded features and clear features and dynamically optimize the channel response of degraded features.

[0034] 4. An unsupervised domain incremental learning method (UDIL) is proposed, which stores the DCROM specific to each domain in the memory bank, and loads the DCROM of the corresponding domain based on the domain discrimination results, thereby achieving zero-forgetting incremental learning. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 The overall architecture of the unsupervised domain incremental learning method guided by global channel priors and the module structure diagram of the dynamic channel response optimization module of the present invention;

[0036] Figure 2 Graph showing the exploration process of inter-domain channel priors in the global channel prior of the present invention;

[0037] Figure 3 Graph showing the exploration process of the intra-domain channel prior in the global channel prior of the present invention;

[0038] Figure 4 This is a diagram of the calculation process of the domain discrimination mechanism of the present invention. DETAILED DESCRIPTION

[0039] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0040] See also Figures 1-4 The present invention provides a technical solution: an unsupervised domain incremental learning method guided by global channel priors, comprising the following steps:

[0041] Step 1: Extract shallow features of the test image through the shallow pre-training layer of the baseline model;

[0042] Through the manifestation of shallow representation space, the same degradation factors have a relatively consistent impact on channel aggregation features, which usually manifests as clustering in feature distribution, such as Figure 2 As shown in the figure, although degradation factors may lead to reduced image clarity and loss of detail, the differences in the distribution of shallow features in different domains mainly depend on the type of degradation and are unrelated to the image content. This phenomenon is called the cross-domain channel prior (CCP), which can serve as a theoretical guide for the domain discrimination mechanism. To verify this view, statistical experiments were conducted on degraded images from the clear domain, foggy domain, and snowy domain.

[0043] Given any set of test images {I1,I2,…,I N}, where I i Represents the i-th test image, N represents the total number of test images, and first uses the VGG16 pre-trained shallow network to extract shallow features of each image Where 224 represents the feature size, 64 represents the number of channels, and the extraction step formula is as follows:

[0044]

[0045] Apply global average pooling technology to Perform feature aggregation by channel to obtain channel aggregation features The formula is as follows:

[0046]

[0047] in, Representation characteristics The value at position (h,w) on channel j, is the global average value of the jth channel;

[0048] Apply t-SNE dimensionality reduction technology to each Projected into two-dimensional space to intuitively represent the feature distribution of the test image, the t-SNE dimensionality reduction process can be expressed as:

[0049]

[0050] in, Represents the two-dimensional feature vector after dimensionality reduction, with a dimension of 1×2, representing the point projected into the two-dimensional space.

[0051] Step 2: Perform global average pooling on shallow features to obtain channel aggregation features;

[0052] Based on the law that the features in the shallow representation space present on the channel response, the channel response of the shallow feature has significant differences in frequency and amplitude, which may be related to the different feature attributes (such as edges, textures, etc.) captured by each channel of the input data, the channel response of the same domain feature is consistent, and the channel response of the different domain feature shows certain difference. This phenomenon is attributed to the introduction of the degradation factor, and is called intro-domain channel prior (ICP), which can be used as a theoretical guide for the channel response optimization module, such as Figure 3 As shown in the figure, the degradation factor changes the channel response distribution of the original clear image, thereby increasing the difficulty of cross-domain feature alignment. To verify the above point, detailed statistical experiment analysis is performed on the image blocks under fog and low light conditions.

[0053] First, randomly select a number of image blocks n from each image I where j∈{1,2,…,K} represents the index of the image block, K is the number of image blocks sampled from the image I n , and the size of each image block is 11x11, which can be represented as:

[0054]

[0055] where x i and y i are the starting coordinates, ensuring that 0≤x i ≤H-11,0≤y i ≤W-11;

[0056] According to the processing steps in steps 1-1 and 1-2, the channel aggregation features of each image block are extracted

[0057] For a group of image blocks, the average value of the channel aggregation features of the group of image blocks is calculated according to the channel to obtain the average channel response The formula is as follows:

[0058]

[0059] Step three, based on the channel aggregation features, the domain discriminant mechanism is used to judge the domain attribute of the test image. If the test image comes from the known domain, the corresponding dynamic channel response optimization module (DCROM) is loaded from the memory bank and integrated into the baseline model for testing. If the test image comes from the unknown domain, switch to the training mode, and use the unpaired clear and test images for unsupervised adversarial training. After training, the generated DCROM is stored in the memory bank.

[0060] In order to alleviate the limitation of domain incremental learning methods that rely on prior knowledge (such as pre-defining the domains that may be encountered in the test phase) to train task identifiers, the present invention designs a domain discrimination mechanism (DDM), such as Figure 4 As shown;

[0061] For any set of test images (e.g. 100), extract the channel aggregation feature matrix according to the preprocessing steps mentioned in the above steps Where 64 represents the number of channels. Subsequently, t-SNE is applied to reduce The channel dimension of

[0062] Will Pre-stored in the dynamic memory storage list (DMSL), which is mainly responsible for storing the feature matrix of each known domain, providing important information for discriminating domain attributes;

[0063] Calculate the center point coordinates of the feature data corresponding to each domain in DMSL Where i represents the domain ID;

[0064] calculate and The distances between L1, L2, ..., L i-1 ;

[0065] Calculate L1, L2, ..., L i-1 The minimum distance L in min ;

[0066] If L min >c, the domain is identified as a new domain, the corresponding feature matrix in DMCL is retained, and the domain ID is set to the list subscript plus 1. Otherwise, the domain is identified as an old domain, the corresponding feature matrix in DMCL is removed, and the domain ID is set to the domain ID corresponding to the minimum distance.

[0067] Step 4: During the training process, the dynamic channel response optimization module is guided by the global channel prior to learn the channel response mapping relationship between the degraded features and the clear features, and dynamically optimize the channel response of the degraded features;

[0068] The main reason why the performance of the source model degrades on degraded images is that the degradation factors change the statistical information of the image, which causes the model's channel response to its shallow features to change, making it significantly different from the response of the clear image, thereby weakening the network's ability to learn features. Therefore, a lightweight dynamic channel response optimization module (DCROM) is designed to optimize the model's channel response to degraded features so that it is closer to the response observed in clear features, such as Figure 1 (c) shows the DCROM structures corresponding to the three baseline models VGG16, ResNet50 and CMT respectively;

[0069] Stage division and initial convolution processing: DCRWM can configure multiple stages, and in each stage Stage i (i = 1, 2, …, m) header, the input features pass through a layer of 3x3 convolution, and the channel dimension is mapped to c*2 i-1 ;

[0070] StarBlock operation: in each stage, the features are processed in turn through n StarBlocks, StarBlock widely uses 1x1 convolution to promote the linear combination of channel responses, thereby enhancing feature interaction and expression ability, 1x1 convolution significantly reduces the huge parameters and computing overhead related to large convolution kernels, in addition, the introduction of star operation helps the fusion of features in different subspaces, and maps the input features to a high-dimensional, nonlinear feature space without widening the network, further improving the model's ability to capture and express complex features, considering the potential redundancy of shallow features, first reduce the channel dimension to reduce feature redundancy while retaining key information, then, the channel dimension is upgraded to enhance the nonlinear expression ability of the network; finally, residual connection is used to reduce the feature loss that may be caused by dimension reduction.

[0071] Further, the global channel prior includes:

[0072] Inter-domain channel prior: channel aggregated features of the same domain tend to cluster, and the cluster center distance between different domains is significantly different, regardless of image content;

[0073] Intra-domain channel prior: channel aggregated features of structurally similar image blocks in the same domain exhibit consistent channel response patterns, while there are significant differences between different domains.

[0074] Further, the domain discrimination mechanism (DDM) includes the following steps:

[0075] 1) Extract the channel aggregated feature matrix of the test image;

[0076] 2) Dimension reduction processing is performed on the channel aggregated feature matrix;

[0077] 3) Calculate the center point coordinates of the dimension-reduced feature matrix and compare them with the feature centers of known domains;

[0078] 4) Determine whether the test image belongs to a known domain according to the minimum distance, if the minimum distance is greater than a preset threshold, it is determined as an unknown domain, if the minimum distance is less than or equal to the preset threshold, it is determined as a known domain, and the corresponding domain ID is assigned.

[0079] Further, the dynamic channel response optimization module (DCROM) includes:

[0080] Stage division and initial convolution processing: at the beginning of each stage, the input features are processed by a 3x3 convolution layer to map the channel dimension to a preset value;

[0081] Star Block operation: in each stage, the features are sequentially processed by multiple Star Block operations, which use a 1x1 convolution to promote the linear combination of channel responses, introduce star operations to fuse features in different subspaces, reduce the channel dimension to reduce redundancy, increase the dimension to enhance nonlinear expression capability, and use residual links to reduce feature loss.

[0082] Further, the unsupervised domain incremental learning method (UDIL) comprises:

[0083] The memory stores DCROM specific to each domain;

[0084] In the training phase, for unknown domain data, the DCROM is optimized using unsupervised adversarial training and stored in the memory;

[0085] In the testing phase, the corresponding DCROM is loaded according to the domain discrimination result and integrated into the baseline model for inference.

[0086] Further, the baseline model includes any one of VGG16, ResNet50 or CMT.

[0087] Further, experiments are conducted on the synthetic degraded data set lmageNet-C, the real fog data set Haze20 and the real low-light data set ExDark, and the method outperforms the baseline model, the unsupervised image restoration method, the unified image restoration method and the test-time adaptive method in terms of Top-1 classification accuracy.

[0088] Further, the dynamic channel response optimization module (DCROM) only updates the weights of the generator and discriminator during training, and the weights of the baseline model remain unchanged.

[0089] Further, the domain discrimination mechanism (DDM) can adapt to an unknown number and any type of domain without providing any prior information.

[0090] Further, the global channel prior is verified by experiments, including channel aggregation feature distribution analysis of degraded images in clear, foggy and snowy domains, and channel response analysis of image blocks in foggy and low-light conditions.

[0091] Embodiment one, Figure 1The framework of the present application is demonstrated, which is different from the previous method. The framework is in test mode at the initial stage and switches to training mode only when an unknown domain is detected. First, any group of test images is extracted by the shallow pre-training layer (SPL) of the baseline model, such as the "Conv2_2" layer of Vgg16 or the "layer1" layer of ResNet50. Then, the shallow features are discriminated by the domain attribute in the domain discrimination mechanism (DDM). If the group of test images comes from the known domain (Known Domain, D k ), the DDM will infer the corresponding domain ID, and the UDIL will load the corresponding DCROM from the memory according to the ID, and then integrate it into the baseline model for testing, as shown in Figure 1 (a). Subsequently, the UDIL continues to remain in the test mode and waits for the subsequent batch of test images. On the contrary, when the images are determined to come from the unknown domain (Unknown Domain, D uk ), the UDIL switches to the training mode and performs unsupervised adversarial training using the unpaired clear and test images, as shown in Figure 1 (b). After the training is completed, the UDIL stores the generated generator obtained by the training as the DCROM in the memory for future use. During the entire training process, the weights of the baseline model remain unchanged, and only the weights of the generator and the discriminator are updated.

[0092] In Example Two, the present application performs comparative experiments on the ImageNet-C synthetic degradation dataset, the Haze20 real fog dataset, and the ExDark real low-light dataset. The Top-1 classification accuracy is used to evaluate the precision of classification. Vgg16, Resnet50, and CMT are used as baseline models for testing. On the ImageNet-C synthetic degradation dataset, the domain incremental learning sequence is: clear → fog → low light → snow. When Vgg16 and Resnet50 are used as baseline models, comparative analysis is performed with the baseline model itself (Source-only), the unsupervised image restoration method (UIR), the unified image restoration method (AIO), and the three test-time adaptive methods (Test-Time Adaption). When CMT is used as a baseline model, comparative analysis is performed with the baseline model itself, the unsupervised image restoration method, and the unified image restoration method. On the Haze20 real fog dataset and the ExDark real low-light dataset, the domain incremental learning sequence is: clear → fog → low light. Vgg16 and Resnet50 are used as baseline models, and comparative analysis is performed with the baseline model itself.

[0093] Table 1 Quantitative comparison results of test results of UIR, All-in-One and Test-time Adaption methods on ImageNet-C synthetic degradation dataset, with Vgg16 and ResNet50 as baseline models

[0094]

[0095] Table 2 Quantitative comparison results of test results of UIR and All-in-One methods on ImageNet-C synthetic degradation dataset, with CMT as baseline model

[0096]

[0097] The results on synthetic degradation data are shown in Tables 1 and 2. Baseline performs differently under different weather conditions. For example, on Vgg16 Baseline, the accuracy can reach 70.4% in clear weather, but only 22.39% in snowy weather. Similar situations also occur on ResNet50, indicating that adverse weather conditions have a significant negative impact on Baseline, highlighting the need for adaptive adjustment of the model. On both Vgg16 and ResNet50 Baselines, the UIR method fails to improve performance, resulting in a decrease in average accuracy to 36.025% and 43.5925%, respectively. This suggests that the UIR method may focus too much on restoring the visual effect of the image, while ignoring the context information of the downstream high-level visual task. The All-in-One image restoration method processes multiple degradation domains through a unified model, optimizing image quality and features simultaneously during model training. The final average accuracy on both Baselines is higher than Source-Only, but the improvement is limited, with only 0.9325 and 0.55, respectively. The TTA method performs adaptive adjustment on the model during the test phase, showing a relatively flexible response ability. However, due to the potential accumulation of errors and catastrophic forgetting problems during the adaptive process, the accuracy of each model in clear weather decreases after 10 rounds of continuous 4-domain transformation, and the performance on the other three domains is uneven, resulting in an average accuracy lower than Source-Only for individual models. In contrast, our method performs well in all weather conditions, far exceeding the baseline model, and outperforms the other three comparison methods in terms of comprehensive performance, demonstrating stronger robustness and adaptability.

[0098] Table 3 Quantitative comparison results of test results of Baseline on Haze20 real fog data set and ExDark real low light data set, with Vgg16 and ResNet50 as baseline models

[0099]

[0100] The results on real degraded data are shown in Table 3. When Vgg16 is used as the baseline model, the performance of the proposed method is relatively small in the foggy condition, while it shows an ideal improvement in the low-light environment, which indicates that Vgg16 may have better generalization in low light. On the contrary, when ResNet50 is used as the baseline model, the proposed method achieves a significant performance improvement of 6.5% in the foggy condition, which may be attributed to the superiority of ResNet50 in processing complex backgrounds and details. However, in the low-light condition, the improvement of ResNet50 is relatively small, only 1.5%, which may be due to the sensitivity of the model to light changes. Although the performances of the two baseline models differ under different degradation conditions, these results still demonstrate the effectiveness of the proposed method on real degraded data.

[0101] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, alternatives and variations can be made to these embodiments without departing from the principles and spirit of the present application, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An unsupervised domain incremental learning method guided by global channel priors, characterized by: The following steps are involved: Step 1: Extract shallow features of the test image through the shallow pre-training layer of the baseline model; Step 2: Perform global average pooling on the shallow features to obtain channel aggregation features; Step 3: Based on the channel aggregation features, the domain attributes of the test image are determined using the domain discrimination mechanism. If the test image comes from a known domain, the corresponding dynamic channel response optimization module (DCROM) is loaded from the memory bank and integrated into the baseline model for testing. If the test image comes from an unknown domain, the training mode is switched to perform unsupervised adversarial training using unpaired clear and test images. After the training is completed, the generated DCROM is stored in the memory bank. Step 4: During the training process, the dynamic channel response optimization module is guided by the global channel prior to learn the channel response mapping relationship between the degraded features and the clear features, and dynamically optimize the channel response of the degraded features.

2. The unsupervised domain incremental learning method guided by global channel prior according to claim 1, characterized in that The global channel prior includes: Inter-domain channel prior: Channel aggregation features in the same domain tend to cluster together, and the distance between cluster centers in different domains is significant, regardless of image content; Intra-domain channel prior: The channel aggregation features of image patches with similar structures in the same domain show consistent channel response patterns, while there are significant differences between different domains.

3. The unsupervised domain incremental learning method guided by global channel prior according to claim 1, characterized in that The domain discrimination mechanism (DDM) includes the following steps: 1) Extract the channel aggregation feature matrix of the test image; 2) performing dimensionality reduction processing on the channel aggregation feature matrix; 3) Calculate the coordinates of the center point of the feature matrix after dimensionality reduction and compare the distance with the feature center of the known domain; 4) Determine whether the test image belongs to a known domain based on the minimum distance. If the minimum distance is greater than the preset threshold, it is determined to be an unknown domain. If the minimum distance is less than or equal to the preset threshold, it is determined to be a known domain and the corresponding domain ID is assigned.

4. The unsupervised domain incremental learning method guided by global channel prior according to claim 1, characterized in that The dynamic channel response optimization module (DCROM) includes: Stage division and initial convolution processing: At the beginning of each stage, the input features pass through a layer of 3x3 convolution to map the channel dimension to the preset value; Star Block operation: In each stage, features are processed in sequence by multiple Star Blocks. 1x1 convolution is used to promote the linear combination of channel responses. Star operations are introduced to fuse features from different subspaces. The channel dimension is first reduced to reduce redundancy and then increased to enhance nonlinear expression capabilities. Residual links are used to reduce feature loss.

5. The unsupervised domain incremental learning method guided by global channel prior according to claim 1, characterized in that The unsupervised domain incremental learning method (UDIL) includes: The memory bank stores each domain-specific DCROM; During the training phase, unsupervised adversarial training is used to optimize DCROM for unknown domain data and store it in the memory bank. During the testing phase, the corresponding DCROM is loaded according to the domain discrimination results and integrated into the baseline model for inference.

6. The unsupervised domain incremental learning method guided by global channel prior according to claim 1, characterized in that The baseline model includes any one of VGG16, ResNet50 or CMT.

7. The unsupervised domain incremental learning method guided by global channel prior according to claim 1, characterized in that Experimental verification is conducted on the synthetic degraded dataset ImageNet-C, the real haze dataset Haze20, and the real low-light dataset ExDark. The proposed method outperforms the baseline model, unsupervised image restoration method, unified image restoration method, and test-time adaptive method in terms of Top-1 classification accuracy.

8. The unsupervised domain incremental learning method guided by global channel prior according to claim 1, characterized in that The dynamic channel response optimization module (DCROM) only updates the weights of the generator and discriminator during training, and the weights of the baseline model remain unchanged.

9. The unsupervised domain incremental learning method guided by global channel prior according to claim 1, characterized in that The proposed domain discrimination mechanism (DDM) can flexibly adapt to an unknown number and arbitrary types of domains without providing any prior information.

10. The unsupervised domain incremental learning method guided by global channel prior according to claim 1, characterized in that The global channel prior is validated through statistical experiments, including channel aggregation feature distribution analysis of degraded images in clear, foggy, and snowy domains, and channel response analysis of image patches under foggy and low-light conditions.

Images