A single-domain generalization target detection method based on cross-domain feature decoupling

Abstract

The application discloses a single-domain generalization target detection method based on cross-domain feature decoupling, which constructs style contrast samples at the target level through a double-flow training structure of parallel input of original images and enhanced images; makes high-quality candidate frames in the original domain and disturbance candidate frames in the enhanced domain jointly participate in instance-level representation learning through cross-domain proposal frame mixing; decouples regional features into semantic subspace and style subspace through a semantic encoder and a style encoder; reduces information coupling between semantic features and style features through orthogonal constraints, domain discrimination constraints and optional mutual information minimization constraints; and finally makes the detection head mainly or only based on semantic features to perform target classification and boundary frame regression. The application is favorable for improving target positioning capability, and is especially suitable for small targets, occluded targets, foggy targets and rainy targets.

Description

Technical Field

[0001] This invention relates to the field of target detection technology, and in particular to a single-domain generalized target detection method based on cross-domain feature decoupling. Background Technology

[0002] Object detection technology is used to locate target objects and identify their categories from images or videos. It is a key technology in autonomous driving environmental perception, traffic monitoring, intelligent security, and robot vision systems. Existing deep learning object detection models typically rely on a large number of labeled samples for training and assume that the data in the training and testing phases follow the same or similar data distribution. However, this assumption is often difficult to hold in real-world open environments. For example, a model may be trained on daytime clear road images but deployed in unknown domains such as nighttime, rainy, foggy, snowy, cross-city roads, low light, strong reflections, image compression, or sensor noise.

[0003] When there are distributional differences between the training and testing domains, object detection models are prone to missed detections, false detections, and bounding box localization errors. A key reason for this is that deep neural networks, trained in a single source domain, easily learn features such as background texture, lighting patterns, weather style, color statistics, and sensor noise—features that are statistically correlated in the source domain but are essentially unrelated to the target category. These features can be referred to as domain-specific style features or non-causal features. When the model is deployed to an unknown target domain, these style features change, leading to a decline in the model's detection performance.

[0004] Existing single-domain generalized object detection methods mainly include data augmentation, style transfer, random convolution, frequency domain perturbation, regularization constraints, feature alignment, and visual-language assistance. Among these, data augmentation methods can expand the appearance variations of training samples, but cannot necessarily eliminate the model's dependence on background style; feature alignment methods may compress class discrimination boundaries, leading to loss of localization information; existing feature decoupling methods suffer from insufficient instance-level constraints, inadequate separation of semantic and style information, and the detection head still being interfered with by style features.

[0005] Therefore, a new single-domain generalized object detection technology is needed to explicitly separate semantic and style information at the instance-level region feature level without accessing the target domain data. This would allow the detection head to rely primarily on the target ontology structure, category semantics, and location-related information, rather than on the source domain background texture, weather disturbances, and lighting style, thereby improving the generalization ability and deployment stability of the object detection model in unknown domains. Summary of the Invention

[0006] The purpose of this invention is to provide a single-domain generalized target detection method based on cross-domain feature decoupling, which is beneficial to improving target localization capabilities, and is especially suitable for small targets, occluded targets, targets in foggy weather and targets in rainy weather.

[0007] To achieve the above objectives, the present invention provides the following solution: A single-domain generalized target detection method based on cross-domain feature decoupling includes the following steps: S1. Obtain the original image of a single source domain. The original image includes the corresponding target category label and target bounding box label. S2. Perform domain perturbation enhancement on the labeled training images to generate enhanced images with the same target semantic annotations as the original images but with different domain styles. The original images and enhanced images constitute a pair of training samples. S3. Input the paired training samples into the preset target detection network. The target detection network includes a shared backbone feature extraction network, a shared region proposal network, a semantic encoder, a style encoder, and a domain classifier. The original image feature map and the enhanced image feature map are obtained through the shared backbone feature extraction network. S4. Generate a set of original image proposal boxes and a set of enhanced image proposal boxes respectively based on the original image feature map and the enhanced image feature map through the shared region proposal network, and perform cross-domain proposal box fusion on the original image proposal box set and the enhanced image proposal box set to obtain a hybrid proposal box set; S5. Based on the hybrid proposal box set, perform region feature extraction on the feature map to obtain instance-level region features; S6. Input the instance-level region features into the semantic encoder and style encoder respectively to obtain semantic features and style features; S7. Apply feature decoupling constraints to semantic features and style features to reduce the correlation between semantic features and style features in the representation space; S8. Perform target classification and bounding box regression based on semantic features, and optimize the target detection network by combining detection loss and feature decoupling loss; S9. In the model inference stage, the image to be detected is input into the optimized target detection network, and the target detection result is output.

[0008] Preferably, in S2, the domain perturbation enhancement of the labeled training image includes at least one of the following: physical prior-based enhancement, illumination perturbation enhancement, atmospheric scattering enhancement, rain and snow particle perturbation enhancement, fogging enhancement, blur enhancement, contrast perturbation enhancement, and style statistical perturbation enhancement.

[0009] Preferably, in S4, cross-domain proposal box blending is performed on the original image proposal box set and the enhanced image proposal box set, as follows:

[0010] in, It is at least one of confidence-based screening, Top-K screening, non-maximum suppression, scale screening, and category-aware screening. For the original image proposal box set, To enhance the image proposal box set.

[0011] Preferably, in S5, region feature extraction is performed on the feature map based on the hybrid proposal box set, specifically including: Based on the hybrid proposal box set, region features are extracted from the original image feature map and the enhanced image feature map respectively, resulting in pairs of original domain instance-level region features and enhanced domain instance-level region features.

[0012] Preferably, in S6, semantic features are used to characterize domain-invariant information related to target category, target outline, target shape, target structure, and target location; style features are used to characterize domain-specific information related to illumination, texture, contrast, weather disturbance, background style, sensor style, or image degradation.

[0013] Preferably, in S7, applying feature decoupling constraints to semantic features and style features includes: Constructing orthogonal decoupling loss :

[0014] in, and The original image is shown below. Normalized semantic features and normalized style features of each sample. and The enhanced image is the first Normalized semantic features and normalized style features of each sample. This represents the number of samples.

[0015] Preferably, S7 also includes a domain discrimination constraint: Style features are input into a domain classifier to determine whether they originate from the original image or the augmented image; a domain discrimination loss is constructed. :

[0016] in, For domain classifiers, As a stylistic feature, These are field labels used to distinguish between the original image and the enhanced image.

[0017] Preferably, in S8, the joint detection loss and feature decoupling loss optimize the target detection network, and the total loss function is as follows:

[0018] in, For classifying losses, For bounding box regression loss, For orthogonal decoupling loss, To determine the loss for the domain, To minimize the loss of mutual information, For semantic consistency loss, This is the loss weight.

[0019] The present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements a single-domain generalized target detection method based on cross-domain feature decoupling as described above.

[0020] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects: (1) This invention constructs a dual-stream training mechanism that runs in parallel with the original image and the enhanced image, allowing the model to simultaneously access different domain style representations of the same target object during the training phase. This enables the construction of style comparison samples at the target level, reducing spurious correlations between the target and specific backgrounds, weather, lighting, or textures. By mixing cross-domain proposal boxes, relatively stable candidate regions in the original image are fused with domain-perturbed candidate regions in the enhanced image, allowing the model to still utilize high-quality candidate boxes for instance-level feature learning in the enhanced domain. This is beneficial for improving target localization capabilities, and is particularly suitable for small targets, occluded targets, foggy targets, and rainy targets.

[0021] (2) This invention explicitly decomposes instance-level region features into semantic and style subspaces using a semantic encoder and a style encoder, and reduces their correlation through orthogonal constraints, structurally reducing the interference of style features on detection decisions. This invention utilizes a domain classifier to supervise the style encoder in absorbing the style differences between the original and enhanced images, concentrating style information in the style subspace, thereby reducing the probability of style information leaking into the semantic subspace. This enables the detection head to perform classification and regression primarily or solely based on semantic features, improving the model's generalization ability to unknown weather, unknown cities, unknown sensors, and unknown image erosion domains without accessing target domain data.

[0022] (3) In the training phase of this invention, decoupling branches can be used to improve robustness, and in the inference phase, auxiliary modules such as enhancement branches, style encoders and domain classifiers can be turned off, thereby maintaining the improvement of generalization performance while avoiding a significant increase in inference cost. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention 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 the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a flowchart illustrating the single-domain generalized target detection method based on cross-domain feature decoupling provided by the present invention. Figure 2 A schematic diagram of the hybrid module structure of the dual-stream training framework and cross-domain proposal box provided by the present invention; Figure 3 This is a schematic diagram of the semantic-style feature decoupling module structure provided by the present invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0027] like Figure 1 As shown, the present invention provides a single-domain generalized target detection method based on cross-domain feature decoupling, comprising the following steps: S1. Obtain the original image of a single source domain. The original image includes the corresponding target category label and target bounding box label. S2. Perform domain perturbation enhancement on the labeled training images to generate enhanced images with the same target semantic annotations as the original images but with different domain styles. The original images and enhanced images constitute a pair of training samples. S3. Input the paired training samples into the preset target detection network. The target detection network includes a shared backbone feature extraction network, a shared region proposal network, a semantic encoder, a style encoder, and a domain classifier. The original image feature map and the enhanced image feature map are obtained through the shared backbone feature extraction network. S4. Generate a set of original image proposal boxes and a set of enhanced image proposal boxes respectively based on the original image feature map and the enhanced image feature map through the shared region proposal network, and perform cross-domain proposal box fusion on the original image proposal box set and the enhanced image proposal box set to obtain a hybrid proposal box set; S5. Based on the hybrid proposal box set, perform region feature extraction on the feature map to obtain instance-level region features; S6. Input the instance-level region features into the semantic encoder and style encoder respectively to obtain semantic features and style features; S7. Apply feature decoupling constraints to semantic features and style features to reduce the correlation between semantic features and style features in the representation space; S8. Perform target classification and bounding box regression based on semantic features, and optimize the target detection network by combining detection loss and feature decoupling loss; S9. In the model inference stage, the image to be detected is input into the optimized target detection network, and the target detection result is output.

[0028] Specifically, the method of the present invention includes: The domain perturbation enhancement in step S2 includes at least one of the following: physical prior-based enhancement, illumination perturbation enhancement, atmospheric scattering enhancement, rain and snow particle perturbation enhancement, fogging enhancement, blurring enhancement, contrast perturbation enhancement, and style statistical perturbation enhancement. Domain perturbation enhancement uses a physical enhancement model to generate enhanced images to simulate open environments such as nighttime, rainy days, foggy days, snowy days, low illumination, high reflectivity, or image erosion.

[0029] The original image and the enhanced image are input into a shared backbone network to obtain feature maps:

[0030] in, For the original image, To enhance the image, To share the backbone feature extraction network, This is the feature map of the original image. To enhance the image feature map.

[0031] The original image feature map and the enhanced image feature map are input into the shared region proposal network, respectively, to obtain:

[0032] in, For the original image proposal box set, To enhance the image proposal box set.

[0033] Cross-domain proposal blending is performed on the original image proposal set and the enhanced image proposal set to obtain a blended proposal set:

[0034] in, This includes confidence screening, Top-K screening, non-maximum suppression, scale screening, category-aware screening, or any combination thereof. This hybrid mechanism allows high-quality candidate regions from the original domain and domain-perturbation candidate regions from the augmented domain to participate in training together, thereby forming style control samples at the target level.

[0035] Region feature extraction from enhanced image feature maps based on a hybrid proposal box set:

[0036] Region feature extraction can also be performed on the original image feature map and the enhanced image feature map separately based on the same set of hybrid proposal boxes:

[0037] in, and These are the instance-level region features corresponding to the original image and the enhanced image, respectively.

[0038] The semantic encoder and style encoder are two parallel nonlinear projectors, respectively satisfying:

[0039] in, For instance-level region features, For semantic encoders, For style encoders, For semantic features, Stylistic characteristics.

[0040] L2 normalization is performed on semantic and stylistic features:

[0041] And construct an orthogonal decoupling loss:

[0042] in, and The original image is shown below. Normalized semantic features and normalized style features of each sample. and The enhanced image is the first Normalized semantic features and normalized style features of each sample. This represents the number of samples.

[0043] The method of this invention also includes style domain discrimination constraints. Style features are input into a domain classifier, which determines whether the style features originate from the original image or the enhanced image, and uses a domain discrimination loss to prompt the style encoder to extract domain-specific style information. The domain discrimination loss is:

[0044] in, For domain classifiers, As a stylistic feature, These are field labels, used to distinguish between the original image and the enhanced image.

[0045] The method of this invention also includes at least one of the following: mutual information minimization constraint, covariance decorrelation constraint, HSIC independence constraint, or adversarial feature separation constraint, to further reduce information leakage between semantic features and style features. For example, cross-covariance penalty can be used:

[0046] It also imposes consistency constraints on the semantic features of the same target instance in the original image and the enhanced image:

[0047] This constraint encourages the same target object to have similar semantic representations under different styles, thereby enhancing the domain invariance of semantic features.

[0048] The joint training loss is:

[0049] in, For classifying losses, For bounding box regression loss, For orthogonal decoupling loss, To determine the loss for the domain, To minimize the loss of mutual information, For semantic consistency loss, This is the loss weight.

[0050] The detection heads used for object classification and bounding box regression only receive semantic features. .

[0051] The object detection network is a two-stage object detection network. Two-stage object detection networks include at least one of Faster R-CNN, Mask R-CNN, and Cascade R-CNN.

[0052] During the inference phase, at least one of the enhanced image generation branch, style encoder, and domain classifier is shut down, and only the backbone feature extraction network, region proposal network, semantic encoder, and detection head are retained to reduce inference computational overhead.

[0053] Example 1 This embodiment provides a single-domain generalized object detection method based on cross-domain feature decoupling. This method uses labeled data from a single source domain as training data and does not require target domain data. The object detection network can use Faster R-CNN as the base detector, or it can be replaced with other two-stage object detection networks.

[0054] During the training phase, for each source domain image First, enhanced images are generated through domain perturbation enhancement. The original image and the enhanced image share the same category labels and bounding box labels. Subsequently, and Simultaneously inputting the shared weights into the backbone network yields the feature map. and Then, a set of corresponding proposal boxes is generated through the shared region proposal network. and The cross-domain proposal box merging module combines two types of proposal boxes into one. and based on Instance-level region feature extraction is performed. The extracted region features are input into the semantic-style decoupling module to obtain semantic features. and style characteristics The two parameters are constrained by orthogonal loss and domain discriminant loss, and finally the detection head outputs classification results and bounding box regression results based on semantic features.

[0055] (1) Enhanced domain perturbation In one embodiment, the domain perturbation enhancement employs a physical enhancement approach, including global illumination perturbation, atmospheric scattering perturbation, and dynamic particle perturbation. Global illumination perturbation is used to simulate nighttime, dusk, low-light, or high-light environments; atmospheric scattering perturbation is used to simulate scattering scenarios in media such as fog, rain, snow, and fog; and dynamic particle perturbation is used to simulate phenomena such as rain streaks, snow particles, motion blur, and particle occlusion.

[0056] In another embodiment, domain perturbation enhancement can also employ random photometric transformation, color jitter, Gaussian blur, JPEG compression, style statistical perturbation, frequency domain perturbation, or generative enhancement. As long as the enhancement step can generate an enhanced image that shares the target semantics with the original image but has a different domain style, it can be used in the two-stream decoupling training framework of this invention.

[0057] (2) Shared feature extraction Original image and image enhancement Input shared backbone network The original image feature map is obtained. and enhanced image feature maps The shared backbone network can be ResNet, ResNeXt, Swin Transformer, ConvNeXt, ViT, CSPDarknet, or other visual feature extraction networks. For two-stage detectors, a Feature Pyramid Network (FPN) can be further incorporated to enhance multi-scale object detection capabilities.

[0058] (3) Cross-domain proposal box hybrid In one embodiment, the original image feature map and the enhanced image feature map are respectively input into a shared RPN to obtain an original image proposal box set and an enhanced image proposal box set. The original image proposal box set typically contains relatively stable, high-quality source domain candidate boxes; the enhanced image proposal box set contains candidate boxes generated under the enhancement style. To enable the detection head to learn cross-domain consistent localization information at the target level, this embodiment mixes the two.

[0059] Preferably, cross-domain proposal box blending may include the following sub-steps: First, the confidence scores of the original set of image proposal boxes and the enhanced set of image proposal boxes are sorted.

[0060] Second, select the candidate boxes that rank highest in their respective categories and construct a union of candidate boxes.

[0061] Third, nonmaximum suppression is performed on the union of candidate boxes to remove highly overlapping and redundant candidate boxes.

[0062] Fourth, supplementary screening is performed based on the target scale, candidate box confidence, category prior, or difficulty level to obtain the final hybrid proposal box set.

[0063] The technical effect of cross-domain proposal box fusion is that the original image branch provides more reliable target localization information, while the augmented image branch provides target appearance information with domain perturbation. The combination of the two can reduce the training instability of the detection model caused by the deterioration of candidate box quality in the augmented domain, while forcing the model to focus on target structures that are stable across styles.

[0064] (4) Region feature extraction After obtaining the hybrid proposal box set, instance-level region features are extracted from the feature map based on RoI Align or RoI Pooling. In one approach, augmentation domain instance-level region features can be extracted on the augmented image feature map. In another approach, paired region features can be extracted simultaneously on both the original and augmented image feature maps.

[0065] In the paired region feature approach, the same hybrid proposal box corresponds to two instance features with different styles but similar target semantics on the original image and the enhanced image, which is beneficial for subsequent semantic consistency constraints and style separation constraints.

[0066] (5) Semantic-style feature decoupling For instance-level region features Two parallel nonlinear encoders are constructed: a semantic encoder and a semantic encoder. and style encoder Semantic encoders are used to extract target category, geometry, contour, shape, component relationships, and positioning-related information; style encoders are used to extract domain-specific information such as background texture, color, lighting, contrast, rain / fog disturbance, and sensor noise.

[0067] To avoid information aliasing between semantic and style features, this embodiment first performs L2 normalization on both and then constructs an orthogonal decoupling loss. When the orthogonal decoupling loss approaches 0, the correlation between semantic and style features in terms of direction decreases, making it difficult for style perturbations to be projected onto the semantic decision direction.

[0068] (6) Style domain determination Using only orthogonal constraints may lead to style branches learning meaningless features. To ensure that the style encoder truly absorbs domain-specific information, this embodiment introduces a domain classifier to determine whether the style features originate from the original image or the enhanced image. When the sample originates from the original image, the domain label can be set to 0; when the sample originates from the enhanced image, the domain label can be set to 1.

[0069] Through this constraint, the style encoder is forced to learn to distinguish the style differences between the original domain and the augmented domain, thereby concentrating domain-specific factors into the style subspace and reducing their risk of entering the semantic subspace.

[0070] (7) Detection head and joint optimization The detection head includes a classification branch and a bounding box regression branch. Unlike traditional detection heads that directly use mixed region features, this embodiment enables the detection head to primarily or solely receive semantic features, thereby blocking the influence of style features on object classification and bounding box regression.

[0071] In practice, joint training can be performed using only classification loss, regression loss, orthogonal decoupling loss, and domain discrimination loss. Alternatively, mutual information minimization loss and semantic consistency loss can be added based on training stability. The loss weights can be set according to the validation set performance or training stability.

[0072] (8) Reasoning stage During the inference phase, the image to be detected does not need to generate an augmented image or be input into a two-stream network. The system can retain only the backbone network, region proposal network, semantic encoder, and detection head. For the input image, feature maps are first extracted, then candidate boxes are generated by the region proposal network, followed by the extraction of region features and the generation of semantic features through the semantic encoder. Finally, the detection head outputs the target category, confidence score, and bounding box location.

[0073] Since the enhancement branch, style encoder, and domain classifier can be turned off during the inference phase, this invention improves the robustness of unknown domain detection without significantly increasing the cost of online inference.

[0074] (9) System Implementation Examples This embodiment also provides a single-domain generalized target detection system based on cross-domain feature decoupling. The system includes a data augmentation module, a shared feature extraction module, a cross-domain proposal box hybridization module, a region feature extraction module, a semantic-style decoupling module, and a detection output module.

[0075] The data augmentation module performs domain perturbation augmentation on single-source-domain labeled training images to obtain augmented images. The shared feature extraction module extracts features from the original and augmented images using shared parameters, resulting in feature maps for the original and augmented images. Figure 2 As shown, the cross-domain proposal box fusion module generates the original image proposal box set and the enhanced image proposal box set respectively, and then fuses them to obtain a hybrid proposal box set. The region feature extraction module extracts instance-level region features based on the hybrid proposal box set. The semantic-style decoupling module maps instance-level region features to semantic and style features and applies feature decoupling constraints to them. The detection output module outputs the target category and target bounding box based on the semantic features.

[0076] In one embodiment, such as Figure 3 As shown, the semantic-style decoupling module includes a semantic encoder, a style encoder, an orthogonal constraint unit, and a domain classifier. The orthogonal constraint unit is used to reduce the directional correlation between semantic features and style features, and the domain classifier is used to determine whether the style features originate from the original image or the enhanced image. The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements a single-domain generalized target detection method based on cross-domain feature decoupling as described above.

[0077] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0078] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, 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 the present invention.

Claims

1. A single-domain generalized target detection method based on cross-domain feature decoupling, characterized in that, Includes the following steps: S1. Obtain the original image of a single source domain, wherein the original image includes the corresponding target category label and target bounding box label; S2. Perform domain perturbation enhancement on the labeled training images to generate enhanced images with the same target semantic annotations as the original images but with different domain styles. The original images and enhanced images constitute a pair of training samples. S3. Input the paired training samples into the preset target detection network. The target detection network includes a shared backbone feature extraction network, a shared region proposal network, a semantic encoder, a style encoder, and a domain classifier. The original image feature map and the enhanced image feature map are obtained through the shared backbone feature extraction network. S4. Generate an original image proposal box set and an enhanced image proposal box set based on the original image feature map and the enhanced image feature map respectively through the shared region proposal network, and perform cross-domain proposal box mixing on the original image proposal box set and the enhanced image proposal box set to obtain a mixed proposal box set. S5. Based on the hybrid proposal box set, perform region feature extraction on the feature map to obtain instance-level region features; S6. Input the instance-level region features into the semantic encoder and style encoder respectively to obtain semantic features and style features; S7. Apply feature decoupling constraints to semantic features and style features to reduce the correlation between semantic features and style features in the representation space; S8. Perform target classification and bounding box regression based on the semantic features, and optimize the target detection network by combining the detection loss and feature decoupling loss; S9. In the model inference stage, the image to be detected is input into the optimized target detection network, and the target detection result is output.

2. The single-domain generalized target detection method based on cross-domain feature decoupling according to claim 1, characterized in that, In S2, the domain perturbation enhancement of the labeled training image includes at least one of the following: physical prior enhancement, illumination perturbation enhancement, atmospheric scattering enhancement, rain and snow particle perturbation enhancement, fogging enhancement, blur enhancement, contrast perturbation enhancement, and style statistical perturbation enhancement.

3. The single-domain generalized target detection method based on cross-domain feature decoupling according to claim 1, characterized in that, In step S4, cross-domain proposal box fusion is performed on the original image proposal box set and the enhanced image proposal box set, as follows: in, It is at least one of confidence-based screening, Top-K screening, non-maximum suppression, scale screening, and category-aware screening. For the original image proposal box set, To enhance the image proposal box set.

4. The single-domain generalized target detection method based on cross-domain feature decoupling according to claim 1, characterized in that, In step S5, region feature extraction is performed on the feature map based on the hybrid proposal box set, specifically including: Based on the hybrid proposal box set, region features are extracted from the original image feature map and the enhanced image feature map respectively, resulting in pairs of original domain instance-level region features and enhanced domain instance-level region features.

5. The single-domain generalized target detection method based on cross-domain feature decoupling according to claim 1, characterized in that, In S6, semantic features are used to characterize target category, target outline, target shape, target structure, and domain-invariant information related to target location; Style features are used to characterize domain-specific information related to lighting, texture, contrast, weather disturbances, background style, sensor style, or image degradation.

6. The single-domain generalized target detection method based on cross-domain feature decoupling according to claim 1, characterized in that, In step S7, applying feature decoupling constraints to semantic features and style features includes: Constructing orthogonal decoupling loss : in, and The original image is shown below. Normalized semantic features and normalized style features of each sample. and The enhanced image is the first Normalized semantic features and normalized style features of each sample. This represents the number of samples.

7. The single-domain generalized target detection method based on cross-domain feature decoupling according to claim 1, characterized in that, The S7 also includes a domain discrimination constraint: Style features are input into a domain classifier to determine whether they originate from the original image or the augmented image; a domain discrimination loss is constructed. : in, For domain classifiers, As a stylistic feature, These are field labels used to distinguish between the original image and the enhanced image.

8. The single-domain generalized target detection method based on cross-domain feature decoupling according to claim 1, characterized in that, In step S8, the joint detection loss and feature decoupling loss optimize the target detection network, and the total loss function is as follows: in, For classifying losses, For bounding box regression loss, For orthogonal decoupling loss, To determine the loss for the domain, To minimize the loss of mutual information, For semantic consistency loss, This is the loss weight.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements a single-domain generalized target detection method based on cross-domain feature decoupling as described in any one of claims 1-8.

Images