A domain generalization semantic segmentation method with causality decoupling and auxiliary clue screening and related device

Abstract

This application discloses a domain generalization semantic segmentation method and related apparatus based on causal decoupling and auxiliary cue selection, relating to the field of computer vision technology. The method includes: extracting multi-scale visual features of the target domain image using a pre-trained visual encoder; decoupling the multi-scale visual features of the target domain into causal branch features and non-causal branch features using a trained learnable mask generator; inputting the causal branch features into a trained causal feature cleansing network to obtain a cleaned causal semantic representation of the target domain; extracting local auxiliary features from the non-causal branch features of the target domain and adaptively fusing them with the cleaned causal semantic representation to obtain fused features of the target domain; and decoding the fused features of the target domain using a trained decoder to obtain the semantic segmentation result of the target domain image. This method improves the segmentation accuracy, boundary quality, and generalization robustness of the model for unknown target domains.

Description

Technical Field

[0001] This application relates to the field of computer vision technology, and in particular to a domain generalization semantic segmentation method and related apparatus for causal decoupling and auxiliary cue selection. Background Technology

[0002] Semantic segmentation, a core task in computer vision, aims to assign semantic category labels to each pixel in an image and is widely used in various intelligent vision fields such as autonomous driving, robot navigation, and medical image analysis. This technology is extensively applied in autonomous intelligent systems such as autonomous driving environmental perception, intelligent robot navigation, and medical image analysis. Traditional semantic segmentation models are typically based on the assumption that training and testing data follow an independent and identically distributed (i.i.d.) distribution. However, in real-world deployment scenarios, the testing environment (target domain) often differs significantly from the training environment (source domain), due to factors such as varying lighting conditions, different weather conditions, changes in sensor types, and shifts in background style. This domain shift problem caused by the distribution difference between the training and testing environments leads to a severe decline in the segmentation performance of pre-trained models in unknown environments. To address this issue, the domain generalization semantic segmentation task has been proposed. Its goal is to enable the model to effectively generalize to unseen target domains using only source domain data for training, without relying on any target domain samples for the training process.

[0003] Existing Domain Generalized Semantic Segmentation (DGSS) methods employ various strategies to enhance model generalization capabilities, including data perturbation enhancement, feature normalization, and invariant representation learning. In recent years, causal representation learning has been introduced into this field. Its core idea is to decouple task-related causal features from task-independent environmental bias features, thereby breaking spurious statistical associations in the source domain and improving the model's adaptability to unknown domains. However, existing methods based on causal decoupling generally suffer from the following drawbacks: useful local visual cues in non-causal branches are completely discarded, environmental biases remain in causal branches, and the decoding stage lacks refined utilization of object edges. Therefore, a domain generalized semantic segmentation technique that can effectively address these issues is urgently needed. Summary of the Invention

[0004] The purpose of this application is to provide a domain generalization semantic segmentation method and related apparatus that combines causal decoupling and auxiliary clue filtering, which can significantly improve the segmentation accuracy, boundary quality and generalization robustness of the model for unknown target domains.

[0005] To achieve the above objectives, this application provides the following solution: Firstly, this application provides a domain-generalized semantic segmentation method for causal decoupling and auxiliary clue filtering, including: A pre-trained visual encoder is used to extract multimodal features from the acquired target domain image to obtain multi-scale visual features of the target domain. A trained learnable mask generator is then used to decouple the multi-scale visual features of the target domain into causal branch features and non-causal branch features of the target domain. The causal branch features of the target domain are input into the trained causal feature purification network to obtain the purified causal semantic representation of the target domain. The target domain local auxiliary features aligned with the purified causal semantic representation of the target domain are selectively extracted from the non-causal branch features of the target domain through a cross-attention mechanism. The target domain local auxiliary features are then adaptively fused with the purified causal semantic representation of the target domain to obtain the target domain fused features. The target domain image and multi-scale visual features of the target domain are processed to obtain target domain boundary enhancement visual features. The cross-modal similarity between the target domain boundary enhancement visual features and the obtained text query embeddings of each category is calculated to obtain the target domain semantic aggregation map. Based on the target domain semantic aggregation map, the target domain fusion features are decoded using a trained decoder to obtain the semantic segmentation result of the target domain image.

[0006] Optionally, the trained learnable mask generator includes a trained global average pooling layer and a trained lightweight multilayer perceptron, the lightweight multilayer perceptron consisting of a linear mapping layer and a sigmoid activation function; a pre-trained visual encoder is used to extract multimodal features from the acquired target domain image to obtain multi-scale visual features of the target domain, and the trained learnable mask generator is used to decouple the multi-scale visual features of the target domain into causal branch features and non-causal branch features of the target domain, specifically including: The target domain image is input into a pre-trained visual encoder, and the visual features output by the intermediate network layers of the pre-trained visual encoder are extracted as multi-scale visual features of the target domain; the pre-trained visual encoder is the visual encoder in the pre-trained contrastive language-image pre-trained model. The global average pooling layer is used to perform global average pooling on the multi-scale visual features of the target domain to obtain the global channel feature vector of the target domain. The global channel feature vector of the target domain is then input into the trained lightweight multilayer perceptron to generate the channel-level mask weight matrix of the target domain. By using the target domain channel-level mask weight matrix, the multi-scale visual features of the target domain are weighted and decomposed to obtain the causal branch features and non-causal branch features of the target domain.

[0007] Optionally, a cross-attention mechanism is used to selectively extract local auxiliary features of the target domain from the non-causal branch features of the target domain, which are aligned with the purified causal semantic representation of the target domain. These local auxiliary features are then adaptively fused with the purified causal semantic representation of the target domain to obtain the target domain fused features, specifically including: After applying a stopping gradient operation to the purified causal semantic representation of the target domain, it is used as a query. The non-causal branch features of the target domain are used as keys and values. The query, keys, and values ​​are mapped to a unified channel dimension through a linear mapping layer, resulting in the target domain query matrix, target domain key matrix, and target domain value matrix. Calculate and normalize the similarity between the target domain query matrix and the target domain key matrix to obtain the target domain attention weight. Multiply the target domain attention weight with the target domain value matrix to extract the local auxiliary features of the target domain. The extracted local auxiliary features of the target domain are concatenated with the purified causal semantic representation of the target domain after the gradient operation is stopped in the channel dimension, and then adaptively fused through a convolutional layer to obtain the target domain fused features.

[0008] Optionally, the target domain image and multi-scale visual features of the target domain are processed to obtain target domain boundary enhancement visual features, and the cross-modal similarity between the target domain boundary enhancement visual features and the obtained text query embeddings of each category is calculated to obtain a target domain semantic aggregation map. Based on the target domain semantic aggregation map, the target domain fusion features are decoded to obtain the semantic segmentation result of the target domain image, specifically including: Boundary priors are extracted from the target domain image to generate high-dimensional structural prior features of the target domain; Dense visual features of the target domain are extracted from the high-dimensional structural prior features of the target domain, and the dense visual features of the target domain are fused with the high-dimensional structural prior features of the target domain to generate visual features that enhance the boundary of the target domain. Obtain text query embeddings for each category; the text query embeddings for each category are extracted from the predefined category text prompts by the text encoder of the contrastive language-image pre-trained model during the training phase; Calculate the cross-modal similarity between the enhanced visual features of the target domain boundary and the text query embeddings of each category to obtain the semantic aggregation map of the target domain; Using the semantic aggregation graph of the target domain as attention guide, the fused features of the target domain are decoded to obtain the semantic segmentation result of the target domain image.

[0009] Optionally, the process for determining the trained learnable mask generator, causal feature cleanup network, and decoder is as follows: Acquire source domain images and pre-built text prompts; text prompts include basic text prompts, context text prompts, and category text prompts; A pre-trained visual encoder is used to process source domain images to obtain multi-scale visual features in the source domain. Then, a current learnable mask generator is used to decouple the multi-scale visual features in the source domain into causal branch features and non-causal branch features in the source domain. A pre-trained text encoder is used to extract text embedding features for text prompts. At the same time, decoupling feature loss is calculated based on the causal branch features and non-causal branch features in the source domain. The text embedding features include basic text embedding features, environmental text embedding features, and category text embedding features. A global environment perturbation dictionary is constructed based on basic text embedding features and environment text embedding features. Perturbation vectors are sampled from the global environment perturbation dictionary. The perturbation vectors are used to perform feature-level modulation on the source domain causal branch features to obtain the perturbation-modulated source domain causal branch features. Based on the source domain causal branch features and the perturbation-modulated source domain causal branch features, the original causal semantic response map and the perturbation-modulated causal semantic response map are obtained. Based on the original causal semantic response map and the perturbation-modulated causal semantic response map, the semantic response consistency loss is calculated. The source domain causal branch features are input into the current causal feature purification network to obtain the causal semantic representation after source domain purification. Based on the causal semantic representation after source domain purification and the non-causal branch features of the source domain, the source domain attention weight is obtained. Based on the source domain attention weight, the source domain local auxiliary features and cross-attention entropy regularization loss are calculated. The source domain local auxiliary features and the source domain purified causal semantic representation are adaptively fused to obtain the source domain fused features. The source domain image and source domain multi-scale visual features are processed to obtain source domain boundary enhancement visual features. The cross-modal similarity between the source domain boundary enhancement visual features and the text query embeddings of each category is calculated to obtain the source domain semantic aggregation map. Based on the source domain semantic aggregation map, the source domain fusion features are decoded using the current decoder to obtain the semantic segmentation result of the source domain image. Based on the semantic segmentation result of the source domain image, the causal branch segmentation loss and the fusion feature segmentation loss are calculated. The current total loss is calculated based on the decoupling feature loss, semantic response consistency loss, cross attention entropy regularization loss, causal branch segmentation loss, fusion feature segmentation loss and the current loss weight; Determine whether the current total loss is less than a preset loss threshold; If so, output the current learnable mask generator, causal feature cleanup network, and decoder as the trained learnable mask generator, causal feature cleanup network, and decoder. If not, the parameters of the current learnable mask generator, causal feature cleansing network, and decoder are updated through backpropagation, and training continues iteratively.

[0010] Optionally, the formula for calculating the total loss is: ; in, Loss due to causal branching; For feature segmentation loss; For decoupling feature loss; This represents the loss of semantic response consistency. The loss is the cross-attention entropy regularization loss; Weights are used to decouple features; Weights for semantic response consistency loss; The weights are used for cross-attention entropy regularization loss.

[0011] Optionally, the formula for calculating the semantic response consistency loss is: ; in, This is the original causal semantic response graph; This is a causal semantic response diagram after perturbation; To stop the gradient operation.

[0012] Optionally, the formula for calculating the cross-attention entropy regularization loss is: ; in, This represents the total number of elements in the source domain attention weight matrix; These are elements in the source domain attention weight matrix.

[0013] In a second aspect, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the domain generalization semantic segmentation method for causal decoupling and auxiliary clue filtering as described above.

[0014] Thirdly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the domain generalization semantic segmentation method for causal decoupling and auxiliary clue filtering as described above.

[0015] Fourthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the domain generalization semantic segmentation method for causal decoupling and auxiliary clue filtering as described above.

[0016] According to the specific embodiments provided in this application, this application has the following technical effects: This application provides a domain generalized semantic segmentation method and related apparatus for causal decoupling and auxiliary cue selection. By using a pre-trained visual encoder to extract multimodal features from the acquired target domain image, multi-scale visual features of the target domain are obtained. Then, a trained learnable mask generator is used to decouple the multi-scale visual features of the target domain into causal branch features and non-causal branch features of the target domain. This solves the problem that existing methods cannot effectively separate cross-domain stable semantics and environmental bias, and achieves the initial decoupling of causal features and non-causal features, laying the foundation for the subsequent selective utilization of useful information in non-causal branches.

[0017] By inputting the causal branch features of the target domain into a trained causal feature purification network, a purified causal semantic representation of the target domain is obtained. This solves the problem that the causal branches in existing causal decoupling methods still retain environmental biases, which limits the generalization ability of the model in unknown domains. It achieves purification of the causal semantic representation and enhances the stability and robustness of the model's cross-domain invariant semantics.

[0018] By selectively extracting local auxiliary features of the target domain aligned with the purified causal semantic representation of the target domain from the non-causal branch features of the target domain through a cross-attention mechanism, and adaptively fusing the local auxiliary features of the target domain with the purified causal semantic representation of the target domain to obtain the target domain fused features, this method solves the problem of existing methods completely discarding non-causal branches and losing local visual cues (such as edges and texture continuity) that are useful for dense prediction tasks. It achieves the preservation and fusion of fine-grained auxiliary information while suppressing harmful domain noise, thereby improving the boundary clarity and detail integrity of the segmentation results.

[0019] By processing the target domain image and multi-scale visual features of the target domain, enhanced visual features of the target domain boundary are obtained. Cross-modal similarity between the enhanced visual features of the target domain boundary and the obtained text query embeddings of each category is calculated to obtain a semantic aggregation map of the target domain. Based on the semantic aggregation map of the target domain, the trained decoder is used to decode the fused features of the target domain to obtain the semantic segmentation result of the target domain image. This solves the problem that existing methods lack fine-grained utilization of object edges in the decoding stage, realizes effective synergy between structural priors and semantic features, and significantly enhances the segmentation precision of high-frequency regions such as object boundaries and thin structures. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is an application environment diagram of a domain generalization semantic segmentation method for causal decoupling and auxiliary clue filtering provided in Embodiment 1 of this application; Figure 2 A flowchart illustrating a domain generalization semantic segmentation method for causal decoupling and auxiliary clue filtering provided in Embodiment 1 of this application; Figure 3 This is a schematic diagram of the first process of the training method provided in Embodiment 2 of this application; Figure 4 This is a second flowchart illustrating the training method provided in Embodiment 2 of this application; Figure 5 This is a detailed structural diagram of step S2 provided in Embodiment 2 of this application; Figure 6 This is a detailed structural diagram of step S3 provided in Embodiment 2 of this application; Figure 7 This is a detailed structural diagram of step S5 provided in Embodiment 2 of this application; Figure 8 This is a detailed structural diagram of step S6 provided in Embodiment 2 of this application; Figure 9 This is a detailed structural diagram of step S6 provided in Embodiment 2 of this application; Figure 10 This is a schematic diagram of the structure of a computer device provided in Embodiment 4 of this application. Detailed Implementation

[0022] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0023] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0024] Example 1 The domain generalization semantic segmentation method for causal decoupling and auxiliary clue filtering provided in this application embodiment can be applied to, for example... Figure 1In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be set up independently, integrated into server 104, or placed in the cloud or on another server. Terminal 102 can send a target domain image to server 104. Server 104 uses a pre-trained visual encoder to extract multi-scale visual features of the target domain image, and uses a trained learnable mask generator to decouple the multi-scale visual features of the target domain into causal branch features and non-causal branch features. The causal branch features are input into a trained causal feature purification network to obtain a purified causal semantic representation of the target domain. Local auxiliary features of the target domain are extracted from the non-causal branch features through a cross-attention mechanism and adaptively fused with the purified causal semantic representation to obtain fused features of the target domain. A trained decoder is used to decode the fused features of the target domain to obtain the semantic segmentation result of the target domain image. Server 104 can feed back the obtained semantic segmentation result of the target domain image to terminal 102.

[0025] The terminal 102 can be, but is not limited to, various desktop computers, laptops, and IoT devices. The server 104 can be implemented using a standalone server or a server cluster consisting of multiple servers, or it can be a cloud server.

[0026] In one exemplary embodiment, such as Figure 2 As shown, a domain-generalized semantic segmentation method for causal decoupling and auxiliary clue filtering is provided. This method is executed by a computer device, specifically by a terminal or server alone, or by both a terminal and a server. In this embodiment, the method is applied to... Figure 1 Taking server 104 as an example, the explanation includes the following steps 201 to 204. Wherein: Step 201: Use a pre-trained visual encoder to extract multimodal features from the acquired target domain image to obtain multi-scale visual features of the target domain, and use a trained learnable mask generator to decouple the multi-scale visual features of the target domain into causal branch features and non-causal branch features of the target domain.

[0027] Step 202: Input the causal branch features of the target domain into the trained causal feature purification network to obtain the purified causal semantic representation of the target domain.

[0028] Step 203: Selectively extract local auxiliary features of the target domain from the non-causal branch features of the target domain through a cross-attention mechanism, which are aligned with the purified causal semantic representation of the target domain. Adaptively fuse the local auxiliary features of the target domain with the purified causal semantic representation of the target domain to obtain the target domain fused features.

[0029] Step 204: Process the target domain image and target domain multi-scale visual features to obtain target domain boundary enhancement visual features, and calculate the cross-modal similarity between the target domain boundary enhancement visual features and the obtained text query embeddings of each category to obtain the target domain semantic aggregation map. Based on the target domain semantic aggregation map, use the trained decoder to decode the target domain fusion features to obtain the semantic segmentation result of the target domain image.

[0030] By implementing steps 201 to 204 as described above, this embodiment has the following beneficial effects: (1) By using a pre-trained visual encoder to extract multimodal features from the acquired target domain image, multi-scale visual features of the target domain are obtained. The multi-scale visual features of the target domain are decoupled into causal branch features and non-causal branch features of the target domain by using a trained learnable mask generator. This solves the problem that existing methods cannot effectively separate cross-domain stable semantics and environmental bias, and realizes the initial decoupling of causal features and non-causal features, laying the foundation for the subsequent selective use of useful information in non-causal branches.

[0031] (2) By inputting the causal branch features of the target domain into the trained causal feature purification network, the causal semantic representation of the target domain after purification is obtained. This solves the problem that the causal branches of the existing causal decoupling methods still have residual environmental bias, which leads to the limited generalization ability of the model in the unknown domain. It realizes the purification of the causal semantic representation and enhances the stability and robustness of the model's cross-domain invariant semantics.

[0032] (3) By selectively extracting local auxiliary features of the target domain from the non-causal branch features of the target domain through the cross attention mechanism, and adaptively fusing the local auxiliary features of the target domain with the causal semantic representation of the target domain after purification, the target domain fused features are obtained. This solves the problem that existing methods completely discard non-causal branches and lose local visual cues (such as edges and texture continuity) that are useful for dense prediction tasks. It achieves the preservation and fusion of fine-grained auxiliary information while suppressing harmful domain noise, and improves the boundary clarity and detail integrity of the segmentation results.

[0033] (4) By processing the target domain image and the target domain multi-scale visual features, the target domain boundary enhancement visual features are obtained, and the cross-modal similarity between the target domain boundary enhancement visual features and the obtained text query embeddings of each category is calculated to obtain the target domain semantic aggregation map. Based on the target domain semantic aggregation map, the target domain fusion features are decoded using the trained decoder to obtain the semantic segmentation result of the target domain image. This solves the problem of the lack of fine utilization of object edges in the decoding stage of existing methods, realizes the effective synergy between structural priors and semantic features, and significantly enhances the segmentation precision of high-frequency regions such as object boundaries and thin structures.

[0034] In summary, this application achieves the purification of domain-invariant causal semantics by combining causal decoupling with auxiliary clue selection, while selectively retaining beneficial local clues in non-causal branches. Furthermore, it introduces edge priors for structural calibration during the decoding stage. As a result, it significantly improves the model's segmentation accuracy, boundary quality, and generalization robustness for unknown target domains under the condition of training only using source domain data.

[0035] Furthermore, the trained learnable mask generator includes a trained global average pooling layer and a trained lightweight multilayer perceptron, which consists of a linear mapping layer and a sigmoid activation function. In step 201, a pre-trained visual encoder is used to extract multimodal features from the acquired target domain image to obtain multi-scale visual features of the target domain. The trained learnable mask generator is then used to decouple the multi-scale visual features of the target domain into causal branch features and non-causal branch features of the target domain, specifically including: Step 2011: Input the target domain image into the pre-trained visual encoder and extract the visual features output by the intermediate network layer of the pre-trained visual encoder as the multi-scale visual features of the target domain; the pre-trained visual encoder is the visual encoder in the pre-trained contrastive language-image pre-trained model.

[0036] Step 2012: Use the trained global average pooling layer to perform global average pooling on the multi-scale visual features of the target domain to obtain the global channel feature vector of the target domain. Then, input the global channel feature vector of the target domain into the trained lightweight multilayer perceptron to generate the channel-level mask weight matrix of the target domain.

[0037] Step 2013: Use the target domain channel-level mask weight matrix to perform weighted decomposition of the target domain multi-scale visual features to obtain the target domain causal branch features and the target domain non-causal branch features; the target domain causal branch features are used to carry cross-domain invariant stable semantics, and the target domain non-causal branch features are used to carry environmental bias and residual details.

[0038] Further, in step 203, a cross-attention mechanism is used to selectively extract local auxiliary features of the target domain from the non-causal branch features of the target domain that are aligned with (i.e., semantically aligned) the purified causal semantic representation of the target domain. These local auxiliary features are then adaptively fused with the purified causal semantic representation of the target domain to obtain the fused features of the target domain, specifically including: Step 2031: Apply a stopping gradient operation to the purified causal semantic representation of the target domain as a query, and use the non-causal branch features of the target domain as keys and values. Map the query, keys, and values ​​to a unified channel dimension through a linear mapping layer to obtain the target domain query matrix, the target domain key matrix, and the target domain value matrix.

[0039] Step 2032: Calculate and normalize the similarity between the target domain query matrix and the target domain key matrix to obtain the target domain attention weight. Multiply the target domain attention weight with the target domain value matrix to extract the local auxiliary features of the target domain.

[0040] Step 2033: The extracted local auxiliary features of the target domain are concatenated with the purified causal semantic representation of the target domain after stopping the gradient operation in the channel dimension, and then adaptively fused through a convolutional layer to obtain the target domain fused features.

[0041] Further, in step 204, the target domain image and multi-scale visual features of the target domain are processed to obtain target domain boundary enhancement visual features, and the cross-modal similarity between the target domain boundary enhancement visual features and the obtained text query embeddings of each category is calculated to obtain a target domain semantic aggregation map. Based on the target domain semantic aggregation map, the target domain fusion features are decoded to obtain the semantic segmentation result of the target domain image, specifically including: Step 2041: Extract the boundary prior of the target domain image to generate high-dimensional structural prior features of the target domain.

[0042] Step 2042: Extract dense visual features of the target domain from the prior features of the high-dimensional structure of the target domain, and fuse the dense visual features of the target domain with the prior features of the high-dimensional structure of the target domain to generate visual features that enhance the boundary of the target domain.

[0043] Step 2043: Obtain the text query embeddings for each category; the text query embeddings for each category are extracted from the preset category text prompts by the text encoder of the contrastive language-image pre-trained model during the training phase.

[0044] Step 2044: Calculate the cross-modal similarity between the target domain boundary enhancement visual features and the text query embeddings of each category to obtain the target domain semantic aggregation map.

[0045] Step 2045: Using the target domain semantic aggregation graph as attention guide, decode the target domain fusion features to obtain the semantic segmentation result of the target domain image.

[0046] Furthermore, the process of determining the trained learnable mask generator, causal feature cleanup network, and decoder is as follows: Step a1: Obtain the source domain image and pre-built text prompts; the text prompts include basic text prompts, environment text prompts, and category text prompts.

[0047] Step a2 involves processing the source domain image using a pre-trained visual encoder to obtain multi-scale visual features in the source domain. The multi-scale visual features in the source domain are then decoupled into causal branch features and non-causal branch features using a current learnable mask generator. Additionally, a pre-trained text encoder is used to extract text embedding features for text prompts. Simultaneously, the decoupling feature loss is calculated based on the causal branch features and non-causal branch features in the source domain. The text embedding features include basic text embedding features, environmental text embedding features, and category text embedding features.

[0048] Step a3: Construct a global environment perturbation dictionary based on basic text embedding features and environment text embedding features. Sample perturbation vectors from the global environment perturbation dictionary. Use the perturbation vectors to perform feature-level modulation on the source domain causal branch features to obtain perturbation-modulated source domain causal branch features. Based on the source domain causal branch features and the perturbation-modulated source domain causal branch features, obtain the original causal semantic response map and the perturbation-modulated causal semantic response map. Calculate the semantic response consistency loss based on the original causal semantic response map and the perturbation-modulated causal semantic response map.

[0049] Step a4: Input the source domain causal branch features into the current causal feature purification network to obtain the purified causal semantic representation of the source domain.

[0050] Step a5: Based on the causal semantic representation after source domain purification and the non-causal branch features of the source domain, obtain the source domain attention weights. Based on the source domain attention weights, calculate the source domain local auxiliary features and the cross-attention entropy regularization loss. Adaptively fuse the source domain local auxiliary features with the source domain purified causal semantic representation to obtain the source domain fused features.

[0051] Step a6: Process the source domain image and source domain multi-scale visual features to obtain source domain boundary enhancement visual features, and calculate the cross-modal similarity between the source domain boundary enhancement visual features and the text query embeddings of each category to obtain the source domain semantic aggregation map. Based on the source domain semantic aggregation map, use the current decoder to decode the source domain fusion features to obtain the semantic segmentation result of the source domain image. Based on the semantic segmentation result of the source domain image, calculate the causal branch segmentation loss and the fusion feature segmentation loss.

[0052] Step a7: Based on the decoupling feature loss, semantic response consistency loss, cross-attention entropy regularization loss, causal branch segmentation loss, fusion feature segmentation loss, and current loss weights, calculate the current total loss.

[0053] Step a8: Determine whether the current total loss is less than the preset loss threshold. If yes, output the current learnable mask generator, causal feature cleanup network, and decoder as the trained learnable mask generator, causal feature cleanup network, and decoder. If no, update the parameters of the current learnable mask generator, causal feature cleanup network, and decoder through backpropagation, and continue iterative training, repeating steps a2-a8.

[0054] Furthermore, the formula for calculating the total loss is as follows: ; in, Loss due to causal branching; For feature segmentation loss; For decoupling feature loss; This represents the loss of semantic response consistency. The loss is the cross-attention entropy regularization loss; Weights are used to decouple features; Weights for semantic response consistency loss; The weights are used for cross-attention entropy regularization loss.

[0055] Furthermore, the formula for calculating the semantic response consistency loss is as follows: ; in, This is the original causal semantic response graph; This is a causal semantic response diagram after perturbation; To stop the gradient operation.

[0056] Furthermore, the formula for calculating the cross-attention entropy regularization loss is as follows: ; in, This represents the total number of elements in the source domain attention weight matrix; These are elements in the source domain attention weight matrix.

[0057] Example 2 The training method provided in the embodiments of this application, such as Figure 3 and Figure 4 As shown, the steps for determining the trained learnable mask generator, causal feature cleanup network, and decoder include: Step S1: Obtain the source domain image and the pre-built text prompt.

[0058] The source domain image is a training image with pre-labeled semantic segmentation labels. Text prompts include basic text prompts, context text prompts, and category text prompts. Basic text prompts are (e.g., "a photo of an object"), context text prompts are (e.g., "a photo of an object in {domain}"), where {domain} can be replaced with descriptions of different domain states such as "snow", "rain", "night", or "fog", and category text prompts are the target category names (e.g., "car", "road", etc.).

[0059] Step S2: Process the source domain image using a pre-trained visual encoder to obtain multi-scale visual features in the source domain, and decouple the multi-scale visual features in the source domain into causal branch features and non-causal branch features in the source domain using the current learnable mask generator; at the same time, extract the text embedding features of the text prompt using a pre-trained text encoder, and calculate the decoupling feature loss based on the causal branch features and non-causal branch features in the source domain.

[0060] Among them, text embedding features include basic text embedding features, contextual text embedding features, and category text embedding features.

[0061] Figure 5 This is a detailed structural diagram of step S2. (See diagram below.) Figure 5 As shown, this step further includes: Step S2.1: Image feature extraction and text feature extraction.

[0062] The input source domain image is denoted as The input is fed into a pre-trained visual encoder to extract features. In this embodiment, the visual encoder uses a ResNet-50 network pre-trained with CLIP to extract visual features from the intermediate layers (e.g., the res4 layer), denoted as... ,in For the number of channels, and This represents the height and width of the feature map.

[0063] The pre-built text prompts are input into the pre-trained CLIP text encoder to obtain the corresponding text embedding features: basic text features. Environmental text features and category text features .

[0064] Step S2.2: Learnable mask generation and feature decomposition.

[0065] This embodiment selects the feature map output from the res4 stage of the backbone network ResNet-50. As a decoupling object, a learnable channel mask generator is constructed, the structure of which is: firstly, for Global average pooling is performed to obtain global channel feature vectors. These vectors are then input into a multilayer perceptron consisting of two linear mapping layers and a sigmoid activation function to generate channel-level mask weight matrices. .

[0066] Using mask matrix Perform eigenvalue decomposition: ; ; in, This indicates element-wise multiplication at the channel level. As a causal branch feature, it carries cross-domain stable semantic information; These are non-causal branching features that carry environmental biases and local auxiliary visual cues.

[0067] Step S2.3: Calculate the decoupling constraint loss.

[0068] To ensure that causal and non-causal branches are statistically independent, a decoupling loss is introduced: ; in, For the mask The L1 sparse regularization penalty term, This is the Hilbert-Schmidt Independence Criterion. The feature mapping function is (first global average pooling, then radial basis function kernel and random Fourier feature approximation). and To balance the weight hyperparameters.

[0069] Step S3: Construct a global environment perturbation dictionary based on basic text embedding features and environment text embedding features; sample perturbation vectors from the global environment perturbation dictionary, and use the perturbation vectors to perform feature-level modulation on the source domain causal branch features to obtain the perturbation-modulated source domain causal branch features; based on the source domain causal branch features and the perturbation-modulated source domain causal branch features, obtain the original causal semantic response map and the perturbation-modulated causal semantic response map, and calculate the semantic response consistency loss.

[0070] Figure 6 This is a detailed structural diagram of step S3. (See diagram below.) Figure 6 As shown, this step further includes: Step S3.1: Construct a class-independent dictionary of environmental perturbations.

[0071] Employing a category-independent anchoring mechanism, the basic text prompt "a photo of an object" is constructed using the neutral word "object," and... The environment text prompt is "a photo of an object in {domain}". Inputting this text prompt into the CLIP text encoder yields the basic text embedding. and the Environmental text embedding .

[0072] Projecting the environmental text embedding onto the orthogonal complement space of the base text embedding yields the environmental perturbation vector: ; Construct a global environment perturbation dictionary Global environment perturbation dictionary .

[0073] Step S3.2: Causal feature modulation based on perturbation dictionary.

[0074] from A perturbation vector is sampled and passed through a lightweight multilayer perceptron to output modulation parameters in the spatial dimension. and , on causal branching characteristics Modulation: ; in This is a hyperbolic tangent activation function used to limit the modulation amplitude.

[0075] Step S3.3: Calculate the semantic response consistency loss.

[0076] Original causal characteristics and modulated causal features Input the prediction head separately to obtain the original semantic response map. Semantic response graph after perturbation .calculate Consistency loss : ; in This indicates that the gradient stop operation is stopped to prevent the gradient of the perturbation branch from interfering with the parameter optimization of the original causal branch.

[0077] Step S4: Input the source domain causal branch features into the current causal feature purification network to obtain the purified causal semantic representation of the source domain.

[0078] The training process of the causal feature sanitization network follows steps S3.1 to S3.3 above to output sanitized causal semantic representations. .

[0079] Step S5: Based on the purified causal semantic representation of the source domain and the non-causal branch features of the source domain, obtain the source domain attention weights; calculate the source domain local auxiliary features and cross-attention entropy regularization loss based on the source domain attention weights; adaptively fuse the source domain local auxiliary features with the purified causal semantic representation of the source domain to obtain the source domain fused features.

[0080] Figure 7 This is a detailed structural diagram of step S5. (See diagram below.) Figure 7 As shown, this step further includes: Step S5.1: Causal-guided cross-attention feature selection.

[0081] Constructing a cross-attention module: using purified causal representation As a query criterion, non-causal branches As keys and values, the input features are first mapped to a uniform channel dimension using a linear mapping layer: ; ; ; Calculate cross-attention to obtain local auxiliary features : ; Among them The gradient stopping operation was used; This is the scaling factor. Let be the channel dimension of the key matrix.

[0082] Step S5.2: Adaptive fusion.

[0083] The causal features and auxiliary features after gradient cessation are concatenated along the channel dimension and then fused using a convolutional layer. ; This is the fusion feature, which is then input into the decoder.

[0084] Step S5.3: Cross-attention entropy regularization loss.

[0085] Apply entropy regularization to the attention weight matrix: ; in For the elements of the attention weight matrix, The total number of elements. This loss leads to a sharp and focused distribution of attention.

[0086] Step S6: Process the source domain image and source domain multi-scale visual features to obtain source domain boundary enhancement visual features; calculate the cross-modal similarity between the source domain boundary enhancement visual features and the text query embeddings of each category to obtain the source domain semantic aggregation map; based on the source domain semantic aggregation map, use the current decoder to decode the source domain fusion features to obtain the semantic segmentation result of the source domain image; calculate the causal branch segmentation loss and the fusion feature segmentation loss based on the semantic segmentation result of the source domain image.

[0087] Figure 8 and Figure 9 This is a detailed structural diagram of step S6. (See diagram below.) Figure 8 and Figure 9 As shown, this step further includes: Step S6.1: Extraction of prior features of image boundaries.

[0088] Given the original input image Spatial edge response maps are extracted using differentiable edge operators (such as the Sobel operator), and then mapped to high-dimensional structural prior features through convolutional layers (including downsampling to align spatial resolution). ,in For the number of channels, and ' represents the height and width of the prior features of the high-dimensional structure.

[0089] Step S6.2: Enhancement of structural and visual feature fusion.

[0090] Select the last layer of features from the multi-scale features obtained by the visual encoder. Dense visual features are obtained through attention pooling. .Will and The concatenation is performed along the channel dimension and then input into a multilayer perceptron to obtain boundary-enhanced visual features: ; Step S6.3: Cross-modal structural semantic joint query and decoding.

[0091] Obtain text query embeddings of various categories generated by CLIP text encoder ( (Total number of categories). Calculate the cross-modal cosine similarity between the boundary-enhanced visual features and the text query embedding to obtain the semantic aggregation graph. : ; The aggregated graph is used as an attention guidance matrix and input into the cross-attention layer of the mask decoder to guide the learnable target query to generate an enhanced semantic query, outputting pixel-level category labels to obtain the semantic segmentation result.

[0092] Step S6.4: Calculate the segmentation loss.

[0093] This embodiment uses two independent monitoring flows: Segmentation loss for independent prediction of causal branches Using only causal branching features The predicted result is obtained through the decoder, and the loss is calculated by comparing it with the true label.

[0094] The final segmentation loss based on fused feature prediction : Using fusion features The final prediction result is obtained through the decoder, and the loss is calculated by comparing it with the true label.

[0095] The segmentation loss is composed of the cross-entropy loss of the query classification. and mask prediction loss It consists of (binary cross-entropy and Dice loss).

[0096] Step S7: Calculate the current total loss based on the decoupling feature loss, semantic response consistency loss, cross-attention entropy regularization loss, causal branch segmentation loss, fusion feature segmentation loss, and the current loss weight.

[0097] The total loss function is: ; in To balance the hyperparameters of each auxiliary loss.

[0098] Step S8: Determine whether the current total loss is less than the preset loss threshold.

[0099] If so, output the current learnable mask generator, causal feature cleanup network, and decoder as the trained model parameters.

[0100] If not, update the parameters of the current learnable mask generator, causal feature cleansing network, and decoder through backpropagation, and return to step S2 to continue iterative training.

[0101] During backpropagation, due to the critical nodes (computation) and (At time) A stopping gradient operation is used, and gradient updates for non-causal branches and auxiliary feature fusion are strictly isolated, thereby protecting the causal backbone network parameters from interference. This embodiment uses stochastic gradient descent (SGD) or the AdamW optimizer to optimize the total loss.

[0102] Example 3 This application also provides an application scenario in which the above-described domain generalization semantic segmentation method for causal decoupling and auxiliary clue filtering is applied. Specifically, the domain generalization semantic segmentation method provided in this embodiment can be applied in a domain generalization semantic segmentation scenario. The domain generalization semantic segmentation scenario includes a semantic segmentation result generation stage for the target domain image and a semantic segmentation result display stage for the target domain image. The semantic segmentation result generation stage is used to perform domain generalization semantic segmentation processing on the acquired target domain image to obtain the semantic segmentation result of the target domain image. The semantic segmentation result display stage is used to display the semantic segmentation result of the target domain image. The domain generalization semantic segmentation method for causal decoupling and auxiliary clue filtering provided in this embodiment belongs to the semantic segmentation result generation stage for the target domain image.

[0103] Example 4 In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 10 As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs stored in the non-volatile storage media. The database stores and processes data. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When the computer program is executed by the processor, it implements a domain generalization semantic segmentation method based on causal decoupling and auxiliary cue filtering.

[0104] Figure 10 The structures shown are merely block diagrams of some structures related to the present application and do not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than shown in the figures, or combine certain components, or have different component arrangements. In an exemplary embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0105] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0106] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0107] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with relevant regulations and be authorized by the owner of the corresponding device.

[0108] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0109] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0110] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0111] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A domain generalization semantic segmentation method decoupled with causal and auxiliary cues screening, characterized in that, The method includes: A pre-trained visual encoder is used to extract multimodal features from the acquired target domain image to obtain multi-scale visual features of the target domain. A trained learnable mask generator is then used to decouple the multi-scale visual features of the target domain into causal branch features and non-causal branch features of the target domain. The causal branch features of the target domain are input into the trained causal feature purification network to obtain the purified causal semantic representation of the target domain. The target domain local auxiliary features aligned with the purified causal semantic representation of the target domain are selectively extracted from the non-causal branch features of the target domain through a cross-attention mechanism. The target domain local auxiliary features are then adaptively fused with the purified causal semantic representation of the target domain to obtain the target domain fused features. The target domain image and multi-scale visual features of the target domain are processed to obtain target domain boundary enhancement visual features. The cross-modal similarity between the target domain boundary enhancement visual features and the obtained text query embeddings of each category is calculated to obtain the target domain semantic aggregation map. Based on the target domain semantic aggregation map, the target domain fusion features are decoded using a trained decoder to obtain the semantic segmentation result of the target domain image.

2. The causally decoupled and auxiliary cue filtered domain generalized semantic segmentation method of claim 1, wherein, The trained learnable mask generator includes a trained global average pooling layer and a trained lightweight multilayer perceptron, which consists of a linear mapping layer and a sigmoid activation function. A pre-trained visual encoder is used to extract multimodal features from the acquired target domain image, resulting in multi-scale visual features of the target domain. The trained learnable mask generator then decouples these multi-scale visual features into causal branch features and non-causal branch features, specifically including: The target domain image is input into a pre-trained visual encoder, and the visual features output by the intermediate network layers of the pre-trained visual encoder are extracted as multi-scale visual features of the target domain; the pre-trained visual encoder is the visual encoder in the pre-trained contrastive language-image pre-trained model. The global average pooling layer is used to perform global average pooling on the multi-scale visual features of the target domain to obtain the global channel feature vector of the target domain. The global channel feature vector of the target domain is then input into the trained lightweight multilayer perceptron to generate the channel-level mask weight matrix of the target domain. By using the target domain channel-level mask weight matrix, the multi-scale visual features of the target domain are weighted and decomposed to obtain the causal branch features and non-causal branch features of the target domain.

3. The causally decoupled and auxiliary cue filtered domain generalized semantic segmentation method of claim 1, wherein, A cross-attention mechanism is used to selectively extract local auxiliary features of the target domain that are aligned with the cleaned causal semantic representation of the target domain from the non-causal branch features of the target domain. These local auxiliary features are then adaptively fused with the cleaned causal semantic representation of the target domain to obtain the target domain fused features, which specifically include: After applying a stopping gradient operation to the purified causal semantic representation of the target domain, it is used as a query. The non-causal branch features of the target domain are used as keys and values. The query, keys, and values ​​are mapped to a unified channel dimension through a linear mapping layer, resulting in the target domain query matrix, target domain key matrix, and target domain value matrix. Calculate and normalize the similarity between the target domain query matrix and the target domain key matrix to obtain the target domain attention weight. Multiply the target domain attention weight with the target domain value matrix to extract the local auxiliary features of the target domain. The extracted local auxiliary features of the target domain are concatenated with the purified causal semantic representation of the target domain after stopping gradient operations in the channel dimension, and then adaptively fused through a convolutional layer to obtain the target domain fused features.

4. The causal decoupling and auxiliary cue screening domain generalization semantic segmentation method according to claim 1, characterized in that, The target domain image and multi-scale visual features of the target domain are processed to obtain target domain boundary enhancement visual features. Cross-modal similarity between the target domain boundary enhancement visual features and the acquired text query embeddings of each category is calculated to obtain a target domain semantic aggregation map. Based on the target domain semantic aggregation map, the target domain fusion features are decoded to obtain the semantic segmentation result of the target domain image, specifically including: Boundary priors are extracted from the target domain image to generate high-dimensional structural prior features of the target domain; Dense visual features of the target domain are extracted from the high-dimensional structural prior features of the target domain, and the dense visual features of the target domain are fused with the high-dimensional structural prior features of the target domain to generate visual features that enhance the boundary of the target domain. Obtain text query embeddings for each category; the text query embeddings for each category are extracted from the predefined category text prompts by the text encoder of the contrastive language-image pre-trained model during the training phase; Calculate the cross-modal similarity between the enhanced visual features of the target domain boundary and the text query embeddings of each category to obtain the semantic aggregation map of the target domain; Using the semantic aggregation graph of the target domain as attention guide, the fused features of the target domain are decoded to obtain the semantic segmentation result of the target domain image.

5. The causally decoupled and auxiliary cue filtered domain generalized semantic segmentation method of claim 1, wherein, The process of determining the trained learnable mask generator, causal feature cleanup network, and decoder is as follows: Acquire source domain images and pre-built text prompts; text prompts include basic text prompts, context text prompts, and category text prompts; A pre-trained visual encoder is used to process source domain images to obtain multi-scale visual features in the source domain. Then, a current learnable mask generator is used to decouple the multi-scale visual features in the source domain into causal branch features and non-causal branch features in the source domain. A pre-trained text encoder is used to extract text embedding features for text prompts. At the same time, decoupling feature loss is calculated based on the causal branch features and non-causal branch features in the source domain. The text embedding features include basic text embedding features, environmental text embedding features, and category text embedding features. A global environment perturbation dictionary is constructed based on basic text embedding features and environment text embedding features. Perturbation vectors are sampled from the global environment perturbation dictionary. The perturbation vectors are used to perform feature-level modulation on the source domain causal branch features to obtain the perturbation-modulated source domain causal branch features. Based on the source domain causal branch features and the perturbation-modulated source domain causal branch features, the original causal semantic response map and the perturbation-modulated causal semantic response map are obtained. Based on the original causal semantic response map and the perturbation-modulated causal semantic response map, the semantic response consistency loss is calculated. The source domain causal branch features are input into the current causal feature purification network to obtain the causal semantic representation after source domain purification. Based on the causal semantic representation after source domain purification and the non-causal branch features of the source domain, the source domain attention weight is obtained. Based on the source domain attention weight, the source domain local auxiliary features and cross-attention entropy regularization loss are calculated. The source domain local auxiliary features and the source domain purified causal semantic representation are adaptively fused to obtain the source domain fused features. The source domain image and source domain multi-scale visual features are processed to obtain source domain boundary enhancement visual features. The cross-modal similarity between the source domain boundary enhancement visual features and the text query embeddings of each category is calculated to obtain the source domain semantic aggregation map. Based on the source domain semantic aggregation map, the source domain fusion features are decoded using the current decoder to obtain the semantic segmentation result of the source domain image. Based on the semantic segmentation result of the source domain image, the causal branch segmentation loss and the fusion feature segmentation loss are calculated. The current total loss is calculated based on the decoupling feature loss, semantic response consistency loss, cross attention entropy regularization loss, causal branch segmentation loss, fusion feature segmentation loss and the current loss weight; Determine whether the current total loss is less than a preset loss threshold; If so, output the current learnable mask generator, causal feature cleanup network, and decoder as the trained learnable mask generator, causal feature cleanup network, and decoder. If not, the parameters of the current learnable mask generator, causal feature cleansing network, and decoder are updated through backpropagation, and training continues iteratively.

6. The causally decoupled and auxiliary cue filtered domain generalized semantic segmentation method of claim 5, wherein, The formula for calculating the total loss is: ; wherein, is a causal branch segmentation loss; is a fused feature segmentation loss; is a decoupled feature loss; is a semantic response consistency loss; is a cross-attention entropy regularization loss; is a decoupled feature loss weight; is a semantic response consistency loss weight; is a cross-attention entropy regularization loss weight.

7. The causally decoupled and auxiliary cue filtered domain generalized semantic segmentation method of claim 6, wherein, The calculation formula of the semantic response consistency loss is: ; wherein, is the original causal semantic response graph; is the perturbed causal semantic response graph; is the stop gradient operation.

8. The causally decoupled and auxiliary cue filtered domain generalized semantic segmentation method of claim 6, wherein, The calculation formula of the cross-attention entropy regularization loss is: ; wherein, is the total number of elements in the source domain attention weight matrix; is an element in the source domain attention weight matrix.

9. A computer device comprising: A memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that the processor executes the computer program to implement the domain generalization semantic segmentation method of causal decoupling and auxiliary clue filtering as described in any one of claims 1-8.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, When executed by a processor, the computer program implements the domain generalization semantic segmentation method of causal decoupling and auxiliary clue filtering as described in any one of claims 1-8.

Images