Single-domain generalization object detection method and system based on evolutionary hint-based adversarial

Abstract

This invention discloses a single-domain generalized object detection method and system based on evolutionary cue adversarial approaches, belonging to the field of object detection technology. The method involves preprocessing the source domain image to obtain an enhanced image; inputting the original image and the enhanced image into a CLIP-ResNet-101 object detection network based on the Faster R-CNN framework to extract dual-branch high-level visual features; generating candidate boxes via a region proposal network; extracting candidate region features using RoIAlign and feeding them into an adversarial domain generation module; constructing text features based on a cross-domain semantic attribute pool and the CLIP text encoder; calculating the domain migration direction and mapping it to a visual offset to generate pseudo-domain features; and generating a continuous pseudo-domain space through an evolutionary learning module in stages, updating the pseudo-domain generator and detector, injecting learnable offsets, training in stages, and dynamically updating the pseudo-domain sample buffer. Detection results are output for images in unknown domains. This invention can effectively improve the model's cross-domain generalization ability and detection stability under single-source domain training conditions, and is suitable for scenarios with strong domain offsets such as lighting and weather conditions.

Description

Technical Field

[0001] This invention relates to the field of target detection technology, and more specifically to a single-domain generalized target detection method and system based on evolutionary cueing adversarial approaches. Background Technology

[0002] Single-domain generalized object detection aims to train a model using only labeled data from a single source domain, enabling it to directly adapt to unknown target domains with significant differences in lighting, weather, and texture distribution. This is a key technology for real-world vision systems such as autonomous driving and intelligent surveillance. In practical deployments, models cannot obtain all target domain data in advance and can only learn within a single source domain. Domain shift leads to catastrophic forgetting, feature confusion, and a sharp drop in detection accuracy, becoming the core bottleneck restricting the technology's practical application. How to improve the model's robustness to continuous domain changes without relying on multi-domain data, while preserving target discrimination features, is a critical problem that urgently needs to be solved in this field.

[0003] Existing single-domain generalization detection methods mainly fall into two categories. The first category is based on feature invariant learning, which decouples visual features into domain-invariant content and domain-related styles through feature decoupling, style separation, and adversarial learning to weaken the influence of domain differences. However, this type of method excessively pursues feature generality, easily discarding domain-specific information with class discriminative power. In scenarios with strong domain shifts, such as extreme lighting and complex textures, the feature representation ability significantly decreases, making it difficult to balance robustness and discriminative power. The second category is based on semantic cues and pseudo-domain expansion, which introduces textual cues using visual language models to generate enhanced samples simulating the target domain to broaden the domain distribution. However, this type of method relies on manually designed discrete cues, cannot model continuous domain changes in real-world scenes, and the cues are static and fixed, lacking dynamic exploration and optimization mechanisms. Insufficient pseudo-domain coverage leads to significant blind spots in generalization.

[0004] In recent years, the academic field has conducted in-depth research on single-domain generalization detection. The CLIP theGap method, proposed at CVPR 2023, was the first to introduce a visual language model into single-domain generalization detection, improving cross-domain feature alignment through textual prompts guiding semantic enhancement and classification loss constraints. The UFR method, proposed at CVPR 2024, constructs a structural causal model from a causal perspective, mitigating data bias and improving domain generalization stability through global-local transformations and causal attention learning. The SRCD method, published in IEEE Transactions on Neural Networks and Learning Systems, further enhances feature generalization capabilities through texture self-enhancement and composite domain semantic reasoning. Despite these advancements, none of these methods have constructed dynamic evolutionary prompts or adversarial pseudo-domain generation mechanisms, and they still struggle to cover continuous domain spaces, exhibiting a significant imbalance in feature learning under complex domain variations.

[0005] Chinese patent CN116563686B discloses a domain-adaptive target detection method based on target domain generalization estimation. This method estimates generalization error by predicting inconsistencies through an auxiliary module and optimizes cross-domain detection performance by combining confidence pseudo-labels. Chinese patent CN118334450A discloses a single-domain generalization method for image data, which improves the model's cross-domain adaptability through feature enhancement and distribution alignment. While these patents improve the domain offset problem to some extent, they still rely on fixed enhancement strategies and static feature constraints. They fail to achieve dynamic evolution of prompts and continuous pseudo-domain construction, exhibiting defects such as pseudo-domain discretization, non-selective distillation, and insufficient exploration of domain changes. Their generalization performance still has significant room for improvement in multi-class strong domain offset scenarios. Summary of the Invention

[0006] To address the aforementioned technical problems, this application discloses a single-domain generalized target detection method and system based on evolutionary cue adversarial approaches. The single-domain generalized target detection method based on evolutionary cue adversarial approaches includes:

[0007] The source image is subjected to basic enhancement processing, including random cropping, Gaussian noise, color dithering, and normalization, to obtain the enhanced image.

[0008] The CLIP-ResNet-101 detection network, based on the Faster R-CNN framework, extracts visual features containing mid-to-high-level semantic information by sharing weights between the original source image and the augmented image inputs. and ;

[0009] use The region proposal network generates candidate proposal boxes. Based on candidate suggestion boxes Through the RoI Align layer, respectively from and Center alignment for extracting candidate region features and ;

[0010] The foreground suggestion box features are input into the adversarial domain generation module. Text features are constructed based on the cross-domain semantic attribute pool and CLIP text encoder. The text domain migration direction is calculated and mapped to visual feature offset. Pseudo-domain features are generated and pseudo-domain features with high detection loss are selected.

[0011] The pseudo-domain features are input into the evolutionary course learning module, and the linear combination of basic attributes and the expansion of offset vectors are performed in stages to generate a continuous pseudo-domain space from easy to difficult.

[0012] An alternating optimization mechanism is adopted. In the pseudo-domain generation stage, the pseudo-domain cue parameters and cross-modal offset mapping function are updated based on adversarial loss, domain offset constraint loss and regularization loss. In the detection training stage, the target detection network parameters are updated based on detection loss. Learnable offsets are injected into Res3 and Res4 layers with fixed probability.

[0013] Training is performed in stages according to the initial task and offset, and the pseudo-domain sample buffer is dynamically updated.

[0014] Input the image of the unknown domain into the trained model, and output the target category label, bounding box coordinates, and confidence score.

[0015] Preferably, the basic enhancement process includes:

[0016] The source domain image is sequentially subjected to fixed-scale random cropping, Gaussian noise addition with preset intensity, color jitter within a fixed range, and mean-variance normalization.

[0017] Enhancement operations only change the statistical distribution of the image appearance, without altering the target size, location, or semantic information of its category;

[0018] The enhanced image maintains the same resolution and channel dimension as the original image, and together they form the input for dual-path feature extraction.

[0019] Preferably, the dual-path feature extraction and proposal box alignment includes:

[0020] A pre-trained CLIP visual encoder is used to extract features from the source and augmented images in parallel, resulting in... and , This is the visual feature map of the original image in the source domain. This is a multi-scale visual feature map of the source domain image after appearance enhancement processing;

[0021] Only with Input RPN to generate suggestion box ;

[0022] Will , and Inputting the RoI Align layer yields foreground proposal box features with a unified dimension. and .

[0023] Preferably, the adversarial domain generation module includes:

[0024] Constructing cross-domain semantic attribute sets ,in, This represents the semantic attributes used to describe cross-domain appearance changes; the set of cross-domain semantic attributes is input into the frozen CLIP text encoder. The basic text features are obtained through the CLIP text encoder. ;

[0025] Based on the aforementioned basic text feature set Through the pseudo-domain generator F Generate pseudo-domain text features:

[0026]

[0027] Embedded from source domain text = ( ) and the pseudo-domain text features Calculate the normalized text domain migration direction:

[0028] =

[0029] And construct ideal pseudo-domain features

[0030]

[0031] in For source domain visual features, Domain migration strength;

[0032] Through cross-modal offset mapping function The text field migration direction The visual offset mapped to the i-th visual feature layer:

[0033] = ( )

[0034] The visual offset is then injected into the source domain visual features. The actual pseudo-domain visual features are obtained:

[0035] = +

[0036] The actual pseudo-domain embedding is obtained through a detection backbone network based on the actual pseudo-domain visual features. ;

[0037] Constructing adversarial losses:

[0038]

[0039] The minimax adversarial optimization strategy is adopted to make the pseudo-domain generator F Search for challenging pseudo-domain text features and apply the cross-modal offset mapping function. The learning process converts the text domain migration direction into the corresponding visual feature offset.

[0040] Preferably, the cross-modal mapping and visual offset injection includes:

[0041] Direction of text field migration Input cross-modal mapping function Output a domain offset vector with the same dimension as the visual features;

[0042] The domain offset vectors are superimposed channel by channel onto the output feature maps of ResNet's Res3 and Res4 layers;

[0043] Offset injection is randomly enabled with a probability of 0.5, introducing domain perturbation without destroying the source domain feature structure, thereby improving the model's adaptability to domain offset.

[0044] Preferably, the pseudo-domain generation method of the evolutionary curriculum learning module is as follows:

[0045] ,

[0046] in Basic prompt features, For attribute weights, A learnable offset vector;

[0047] Phase 1 Only optimize The pseudo-domain is generated by a linear combination of basic attributes;

[0048] Phase Two Activation This expands the feature space and generates complex pseudo-domains, enabling learning from easy to difficult courses.

[0049] Preferably, the alternating optimization includes two steps:

[0050] Step 1: Pseudo-domain generation optimization ,in To generate loss for adversarial domains, For complexity regularization loss;

[0051] Step 2: Detector optimization ,in For source domain feature maps, It is a pseudo-domain transformation function. For labeling.

[0052] Preferably, overall model optimization includes:

[0053] The detection classification loss, bounding box regression loss, adversarial pseudo-domain generation loss, and pseudo-domain regularization loss are weighted and fused to obtain the overall optimization objective.

[0054] During the detector optimization phase, the target detection network parameters are updated based on the detection classification loss and bounding box regression loss.

[0055] During the pseudo-domain generation optimization stage, the pseudo-domain generator, cross-modal offset mapping function, and learnable visual offset parameters are updated based on the adversarial pseudo-domain generation loss and regularization loss.

[0056] By alternately executing the detector optimization phase and the pseudo-domain generation optimization phase, the target detection network gradually adapts to the continuously generated pseudo-domain features, which range from easy to difficult.

[0057] Preferably, the phased training and offset update include:

[0058] In the first stage, the learnable residual perturbation vector is frozen or set to zero. Based solely on the basic text feature set Optimize attribute combination weights Generate initial pseudo-domain text features located in the basic attribute space;

[0059] In the second stage, a learnable residual perturbation vector is enabled based on the initial pseudo-domain text features obtained in the first stage. This expands the search range of pseudo-domain text features and generates more complex and challenging pseudo-domains.

[0060] For each candidate pseudo-domain, calculate its adversarial loss, and select the top K highly challenging pseudo-domains based on the loss magnitude to participate in subsequent optimization;

[0061] Based on the loss gain before and after enabling the residual perturbation vector To determine whether the residual disturbance is effective, when Greater than the preset threshold If the corresponding pseudo-domain candidate is retained, its weight in training should be reduced or it should be removed.

[0062] During the pseudo-domain candidate update process, orthogonal regularization constraints are used to maintain the diversity among different pseudo-domain candidates in order to stabilize the expansion process of the continuous pseudo-domain space.

[0063] The single-domain generalized target detection system based on evolutionary cueing adversarial mechanisms includes:

[0064] The data preprocessing module performs random cropping, Gaussian noise reduction, color dithering, and normalization enhancement on the source domain image;

[0065] The dual-path feature extraction and candidate box alignment module is used to extract source domain visual features and enhanced visual features from the CLIP-ResNet-101 detection network based on the Faster R-CNN framework, which inputs the source domain original image and the enhanced image with shared weights. The module generates candidate proposal boxes based on the source domain visual features and then aligns the source domain visual features and enhanced visual features with candidate region features through the RoI Align operation.

[0066] The adversarial domain generation module constructs text attribute embeddings, calculates domain migration directions, generates pseudo-domain features, and performs adversarial optimization.

[0067] The evolutionary curriculum learning module generates linear combination pseudo-domains and extended offset pseudo-domains in stages, and dynamically adjusts the difficulty of the pseudo-domains.

[0068] The training module is optimized alternately, and pseudo-domain generation optimization and detector optimization are performed alternately, with learnable domain offsets injected into the intermediate layer.

[0069] The phased scheduling and pseudo-domain candidate region management module executes the initial task and offset training task, and dynamically updates the pseudo-domain sample set.

[0070] The cross-domain detection inference module performs forward inference on images in unknown domains and outputs the category, bounding box coordinates, and confidence score.

[0071] Compared with the prior art, the technical solution of this application has the following technical effects:

[0072] This invention constructs a detection framework that combines adversarial domain generation and evolutionary curriculum learning. It can automatically uncover potential domain differences and generalization blind spots under single source domain training conditions, enabling the model to continuously engage with diverse domain distributions during training. This effectively improves the model's adaptability to unknown target domains and its cross-domain generalization stability, ensuring that the model maintains reliable feature extraction and target recognition capabilities in various domain shift scenarios, and providing stable support for target detection in complex environments.

[0073] This invention employs a cross-modal mapping mechanism of text prompts and visual features, transforming the domain migration direction at the semantic level into feature shifts in the visual space. This enables accurate modeling of continuously changing domain distributions, expands the domain coverage that the model can learn, and strengthens the learning and retention of category-specific discriminative features. It avoids feature confusion and foreground / background interference, improves the model's accuracy and discriminative ability in representing target features, and ensures the consistency and reliability of detection results.

[0074] This invention employs a phased evolutionary learning strategy to dynamically adjust the difficulty of pseudo-domain generation, enabling the model to gradually adapt to changes in complex domains from simple to complex. This significantly improves the stability and convergence efficiency of the training process, while reducing overfitting and domain bias of the model to single source domain data. It stably retains learned knowledge while continuously learning new domain features, balancing the model's plasticity and stability, and ensuring a smooth and reliable multi-step offset training process.

[0075] This invention employs an alternating optimization and intermediate layer learnable offset injection mechanism to achieve collaborative iteration of pseudo-domain generation and detector optimization without relying on multi-source domain data and complex alignment operations. This simplifies the model training and deployment process, reduces computational and storage overhead, and improves the versatility and engineering practicality of the overall framework. The model can efficiently complete end-to-end single-domain generalized target detection, meeting the deployment needs of various practical scenarios.

[0076] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the preferred embodiments of this application are described in detail below with reference to the accompanying drawings.

[0077] The above and other objects, advantages and features of this application will become more apparent to those skilled in the art from the following detailed description of specific embodiments in conjunction with the accompanying drawings. Attached Figure Description

[0078] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In all drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0079] Based on the description of the figures and their corresponding technical content in the document, the titles of the figures are as follows:

[0080] Figure 1 Flowchart of a single-domain generalized target detection method based on evolutionary cueing adversarial approaches;

[0081] Figure 2 A schematic diagram of a cross-modal transfer architecture for evolutionary cue adversarial and alternating optimization training;

[0082] Figure 3 Graph showing the performance trend of the model under different domain migration strength coefficients λ;

[0083] Figure 4 A bar chart comparing the cross-domain detection performance of the model under different course learning parameters τ;

[0084] Figure 5 Module architecture diagram of a single-domain generalized target detection system based on evolutionary cueing adversarial mechanisms. Detailed Implementation

[0085] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, 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, not all embodiments. In the following description, specific details such as specific configurations and components are provided merely to help fully understand the embodiments of this application. Therefore, those skilled in the art should understand that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. In addition, for clarity and brevity, descriptions of known functions and structures are omitted in the embodiments.

[0086] It should be understood that the phrase "an embodiment" or "this embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "an embodiment" or "this embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments.

[0087] Furthermore, reference numerals and / or letters may be repeated in different examples within this application. Such repetition is for the purpose of simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or settings discussed.

[0088] In this article, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, B exists alone, and A and B exist simultaneously. The term " / and" in this article describes another type of relationship between related objects, indicating that two relationships can exist. For example, A / and B can mean: A exists alone, and A and B exist alone. In addition, the character " / " in this article generally indicates that the related objects before and after it are in an "or" relationship.

[0089] In this article, the term "at least one" is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, "at least one of A and B" can mean: A exists alone, A and B exist simultaneously, or B exists alone.

[0090] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion.

[0091] Example 1

[0092] This embodiment mainly describes a single-domain generalized target detection method based on evolutionary cueing adversarial approaches, such as... Figure 1 As shown, it specifically includes:

[0093] The source image is subjected to basic enhancement processing, including random cropping, Gaussian noise, color dithering, and normalization, to obtain the enhanced image.

[0094] Faster R-CNN and ResNet-101 networks, which share weights between the original source image and the augmented image, extract multi-scale feature maps respectively. and ;

[0095] use The region proposal network generates candidate proposal boxes. and through the RoI Align layer from and Center alignment for foreground suggestion box feature extraction and ;

[0096] Input the foreground suggestion box features into the adversarial domain generation module, construct text features based on the cross-domain semantic attribute pool and CLIP text encoder, calculate the text domain migration direction and map it to visual feature offset, generate pseudo-domain features and mine the model's generalization blind spot.

[0097] The pseudo-domain features are input into the evolutionary course learning module, and the linear combination of basic attributes and the expansion of offset vectors are performed in stages to generate a continuous pseudo-domain space from easy to difficult.

[0098] An alternating optimization mechanism is adopted. First, the detector is frozen to optimize the pseudo-domain generator and cross-modal mapping function against the loss and regularization loss. Then, the pseudo-domain generator is frozen to optimize the detector with the detection loss. Learnable biases are injected with fixed probability in Res3 and Res4 layers.

[0099] Multi-step training is performed in stages according to the initial task and offset training task, and the pseudo-domain sample buffer is dynamically updated.

[0100] Input the image of the unknown domain into the trained model, and output the target category label, bounding box coordinates, and confidence score.

[0101] Furthermore, for the input source domain image Perform basic enhancement processing; the enhancement function is denoted as... The process involves sequentially performing fixed-scale random cropping, adding Gaussian noise with a preset standard deviation, color jittering within a fixed channel range, and zero-mean unit variance normalization to obtain an enhanced image with the same resolution and number of channels as the original image. The original image of the source domain. With enhanced image The Faster R-CNN framework with synchronous input and shared weights uses ResNet-101 as the backbone network for forward feature extraction, and outputs multi-scale visual feature maps of the source domain. With augmented domain multi-scale visual feature maps Both have the same number of channels, height, and width. Only... Input the Region Proposal Network (RPN) to generate a set of target candidate suggestion boxes. .Will , and The input RoI Align layer performs bilinear interpolation and pooling alignment according to a fixed output size to obtain foreground proposal box features with uniform dimensions. and ,satisfy:

[0102]

[0103]

[0104] in This indicates the RoI Align region feature extraction operation based on candidate proposal boxes, with a fixed and consistent output feature dimension. Characterize the foreground proposal box features of the original image. The foreground proposal box features of the enhanced image are represented, and the two sets of features are used together as input to the subsequent adversarial domain generation module.

[0105] Furthermore, the Adversarial Domain Generation Module (ADGM) specifically involves: constructing a cross-domain semantic attribute set. Each attribute in the attribute set is input into the CLIP text encoder to obtain a set of basic text feature vectors. Embedded with source domain text Based on this, calculate the domain migration direction in the text space. Based on source domain visual features Constructing an ideal pseudo-domain feature representation:

[0106]

[0107] In the formula, The visual features output from the source domain image via the backbone network are the visual features. This is the domain migration strength coefficient, used to control the pseudo-domain offset magnitude; it is obtained through the cross-modal mapping function. Will The feature map is converted into a visual space feature offset, and then injected with the source domain features to obtain a pseudo-domain feature map, which in turn yields the pseudo-domain visual features of the actual model output. Constructing an adversarial loss function to achieve highly challenging pseudo-domain mining: ,use Optimize the adversarial capabilities, among which For cross-modal mapping functions, It is a pseudo-domain generator that identifies unfit domains by maximizing feature differences and achieves domain transfer learning by minimizing mapping errors.

[0108] Furthermore, the Evolutionary Learning Module (ECL) specifically includes: a pseudo-domain generator. The model is a dynamic difficulty scheduling structure, and the pseudo-domain representation is generated jointly by weighted combination and offset expansion, specifically in the following form: In the formula, The text feature vector corresponding to the basic cue pool. These are learnable attribute weights, used to achieve dynamic weighted combinations of basic attributes. These are learnable offset vectors used to expand the pseudo-domain feature space and improve complexity. The training process is strictly divided into two stages; the first stage will... Set to 0 to optimize only attribute weights. The pseudo-domain is composed of a linear combination of basic attributes; the second phase is enabled. This expands the feature space and generates a highly complex pseudo-domain, which is then used to calculate the difference through loss. Measure the increase in complexity to achieve a gradual learning process from easy to difficult.

[0109] Furthermore, the alternating optimization mechanism and intermediate layer offset injection are used to collaboratively update the pseudo-domain generation structure and detector parameters in the overall model.

[0110] Step 1: Pseudo-domain generation. Freeze the detector parameters and focus on generating pseudo-domains. The aim is to produce pseudo-domains that the detector is not adapted to, while avoiding generating overly complex pseudo-domains. In this step, we jointly optimize the transformation function T. and pseudo-domain generator F Its loss function is: ,in, It is the adversarial loss in ADGM, used to train the pseudo-domain generator F. , It is a regularization loss used to prevent the generated pseudo-domains from becoming too complex and to maintain the stability of training.

[0111] The second step is detector optimization. The pseudo-domain generator parameters are frozen, and the detection model is trained based on pseudo-domain features. The optimization objective is: ,in For the input image, For source domain feature maps, It is a pseudo-domain transformation function. For labeling, It includes classification loss and bounding box regression loss. Furthermore, learnable offsets are introduced after the Res3 and Res4 feature layers of ResNet, randomly superimposed onto the feature map with a probability of 0.5, improving the model's adaptability to domain offsets. By generating pseudo-domains and combining ADGM and ECL, an alternating two-step optimization process is used to train the object detector.

[0112] During training, the model undergoes two-step optimization, alternating between steps. In the first step, we optimize the pseudo-domain generation part to ensure the generator can recognize and generate new pseudo-domains. In the second step, we optimize the detector to adapt to the generated pseudo-domain features, thereby improving detection performance. Furthermore, to enhance the model's adaptability to different target domains, learnable offsets are introduced after the Res3 and Res4 layers. These offsets are applied to the network with a probability of 0.5 to enhance the detector's robustness. These offsets are used to improve target detection performance. The entire framework process is as follows: Figure 2 As shown;

[0113] Furthermore, model training is divided into an initial task phase and an offset training phase. In the initial task phase, the base detector is trained using the initial class dataset from the source domain. Only the detection loss constraint is applied, and after initial model convergence, adversarial feature mining and pseudo-domain template construction are initiated. In the offset training phase, pseudo-domain enhancement training, adversarial feature mining, pseudo-domain template updates, and classifier weight fusion are performed iteratively to gradually expand the domain generalization coverage. After each round of offset training, the pseudo-domain sample buffer is dynamically updated based on the newly added class information and the current pseudo-domain distribution. Redundant samples are removed, and highly representative pseudo-domain samples are added to maintain a balanced sample distribution in the buffer, reducing the model's overfitting to the source domain and maintaining domain generalization stability.

[0114] Furthermore, after model training, end-to-end forward inference is performed on the unknown domain image. The unknown domain image is input into the backbone network to complete multi-scale feature extraction. After pseudo-domain feature enhancement, adversarial calibration, and cross-modal alignment, the result is input into the detection head. The model directly outputs the target category label, bounding box coordinates, and confidence score. It does not perform target domain fine-tuning, does not rely on target domain annotation, and does not perform additional domain adaptation throughout the process, achieving strong robust single-domain generalized target detection under single source domain training.

[0115] This detailed implementation describes how adversarial domain generation and evolutionary curriculum learning work together to automatically uncover continuous domain variations and generalization blind spots under single-source domain training. It leverages cross-modal mapping to achieve accurate domain transfer from textual cues to visual features, constructing a comprehensive pseudo-domain space. A phased, progressive learning process stabilizes the training process, while alternating optimization and intermediate layer offset injection balance feature robustness and discriminative power. Dynamic buffer updates reduce source domain overfitting, enabling the model to possess stable and efficient cross-domain generalization detection capabilities even without target domain data.

[0116] Based on Example 1, this example details the technical verification of this application. The core challenge of existing single-domain generalized object detection lies in how to construct a robust detection model for unknown target domains with only single source domain labeled data. Existing methods mostly rely on static cues, discrete pseudo-domains, or fixed enhancement strategies, making it difficult to model continuous and complex domain distribution changes in real-world scenes, resulting in a significant decrease in detection performance under strong domain shift scenarios. The single-domain generalized object detection method based on evolutionary cue adversarial learning proposed in this application constructs a continuous pseudo-domain space through adversarial pseudo-domain generation and evolutionary curriculum learning, achieving accurate domain transfer from text cues to visual features through cross-modal mapping. This application conducts systematic comparative experiments on a standard single-domain generalized object detection dataset, using Daytime-clear as the source domain, and tests it on various unknown target domains to comprehensively evaluate its cross-domain generalization performance.

[0117] This application selects several mainstream single-domain generalization detection methods as baselines and uses mAP as the quantitative metric to conduct comparative experiments on public datasets to verify the performance advantages of the proposed method. The comparison results are shown in the table below.

[0118] Faster R-CNN 54.7 34.0 30.5 14.0 32.2 27.2 SW 50.6 33.4 26.3 13.7 30.8 26.1 IBN-Net 49.7 32.1 26.1 14.3 29.6 25.5 IterNorm 43.9 29.6 22.8 12.6 28.4 23.4 ISW 51.3 33.2 25.9 14.1 31.8 26.3 S-DGOD 56.1 36.6 28.2 16.6 33.5 28.8 CLIP the Gap 51.3 36.9 32.3 18.7 38.5 31.6 Prompt-Driven 53.6 38.5 33.7 19.2 39.1 32.6 OA-DG 55.8 38.0 33.9 16.8 38.3 31.8 UFR 58.6 40.8 33.2 19.2 39.6 33.2 TGFD 52.4 36.8 31.3 16.7 38.2 30.8 Ours 55.5 41.9 35.2 20.6 41.4 34.8

[0119] The table above presents the mAP comparison results between the proposed method and existing mainstream methods in different target domain scenarios. In the single source domain Daytime-clear scenario, the mAP of the proposed method is 55.5%, which is comparable to most baseline methods, indicating that it does not cause significant loss in source domain performance and can achieve improved cross-domain generalization performance without compromising source domain detection capabilities. In all unknown target domain scenarios, the proposed method achieves significant performance improvements. Specifically, the mAP in the Night-sunny domain reaches 41.9%, which is 1.1 percentage points higher than the second-best UFR method; the mAP in the Dusk-rainy domain is 35.2%, which is 1.5 percentage points higher than the Prompt-Driven method; the mAP in the Night-rainy domain is 20.6%, which is 1.4 percentage points higher than the UFR method; and the mAP in the Daytime-foggy domain is 41.4%, which is 1.8 percentage points higher than the UFR method. In terms of average performance, the method in this application achieves an average mAP of 34.8%, which is 1.6 percentage points higher than the second-best UFR method and comprehensively outperforms other baseline methods. This fully verifies the cross-domain generalization ability and universality of the method in strong domain offset scenarios and proves the effectiveness of the adversarial pseudo-domain generation and evolutionary curriculum learning mechanism.

[0120] To further investigate the impact of key hyperparameters on model performance, this application conducted a systematic sensitivity analysis on the domain migration strength coefficient λ, and the results are as follows: Figure 3 As shown, Figure 3 The figure shows the mAP variation curves of the model in two typical target domains, Night-Sunny and Daytime-Foggy, under different values ​​of λ. As can be seen from the figure, as λ gradually increases from 0.1 to 0.5, the mAP of both target domains shows a continuous upward trend. When λ=0.5, the mAP of the Night-Sunny domain reaches a peak of 41.9%, and the mAP of the Daytime-Foggy domain also reaches 41.4%. However, when λ continues to increase to 0.7 and 1.0, the mAP of both target domains decreases to varying degrees. This result indicates that when the value of λ is too small, the offset amplitude of the generated pseudo-domain is insufficient, failing to effectively cover the domain offset distribution in the real scene, thus limiting the model's generalization ability to unknown domains. Conversely, when the value of λ is too large, the generated pseudo-domain differs too much from the source domain distribution, easily introducing irrelevant noise features, causing the model to learn patterns that lack generalization ability, thereby reducing detection performance. Therefore, λ=0.5 is the optimal value for the method in this application, which can maintain the correlation with the source domain features while ensuring the diversity of pseudo-domains, and achieve the best cross-domain generalization effect.

[0121] This application also conducted a sensitivity analysis on the control parameter τ during the course learning phase, and the results are as follows: Figure 4 As shown, Figure 4 The figure shows the mAP variation curves of the model in the Night-Sunny and Daytime-Foggy target domains for different values ​​of τ. As can be seen from the figure, as τ increases from 0.02 to 0.05, the mAP of both target domains significantly improves. When τ = 0.05, the mAP of the Night-Sunny domain reaches a peak of 41.9%, and the mAP of the Daytime-Foggy domain also reaches 41.4%. When τ continues to increase to 0.1 and 0.2, the mAP of both target domains gradually decreases. This result indicates that when τ is too small, the learning phase transitions too early, and the model enters a high-complexity stage before fully learning the pseudo-domain features of the basic attribute combinations, easily leading to training instability and insufficient pseudo-domain learning. Conversely, when τ is too large, the model remains in the basic stage for too long, failing to introduce learnable offsets to expand the pseudo-domain space in time, resulting in insufficient pseudo-domain diversity and limited generalization ability. Therefore, τ=0.05 is the optimal value for the method in this application, which can ensure a smooth transition from easy to difficult learning, and enable the model to fully explore the features of high-complexity pseudo-domains while training stably, thereby effectively improving cross-domain generalization performance.

[0122] This embodiment achieves significant performance improvements in multiple strong domain shift scenarios through collaborative optimization of adversarial pseudo-domain generation and curriculum learning. All indicators are superior to existing mainstream solutions, with an average mAP of 34.8%. At the same time, the parameter sensitivity is controllable, training is stable, and it can effectively enhance the model's cross-domain generalization ability and engineering practicality.

[0123] Example 2 describes in detail a single-domain generalized target detection system based on evolutionary cue adversarial mechanisms, such as... Figure 5 As shown, it includes:

[0124] The data preprocessing module performs random cropping, Gaussian noise reduction, color dithering, and normalization enhancement on the source domain image;

[0125] The dual-path feature extraction module extracts multi-scale features from two branches based on shared-weight Faster R-CNN and ResNet-101 and completes RoI Alignment.

[0126] The adversarial domain generation module constructs text attribute embeddings, calculates domain migration directions, generates pseudo-domain features, and performs adversarial optimization.

[0127] The evolutionary curriculum learning module generates linear combination pseudo-domains and extended offset pseudo-domains in stages, and dynamically adjusts the difficulty of the pseudo-domains.

[0128] The training module is optimized alternately, and pseudo-domain generation optimization and detector optimization are performed alternately, with learnable domain offsets injected into the intermediate layer.

[0129] The phased scheduling and buffer management module executes the initial task and offset training task, and dynamically updates the pseudo-domain sample set.

[0130] The cross-domain detection inference module performs forward inference on images in unknown domains and outputs the category, bounding box coordinates, and confidence score.

[0131] Furthermore, the data preprocessing module performs basic enhancement processing on the source domain image using a fixed workflow, including:

[0132] Fixed-scale random cropping, preset intensity Gaussian noise addition, fixed-range color jitter, and global mean-variance normalization.

[0133] The enhancement only changes the statistical distribution of the image appearance, without changing the target size, location, or category semantics. This ensures that the enhanced image and the original image have completely consistent resolution and channel dimensions, providing a standard input for subsequent dual-path feature extraction.

[0134] Furthermore, the dual-path feature extraction module constructs a dual-branch parallel feature extraction structure based on shared-weight Faster R-CNN + ResNet-101:

[0135] The input is the original source domain image, and the output is a multi-scale feature map. ;

[0136] Input is a base-enhanced image, output is a multi-scale feature map ;

[0137] The two feature paths share the same set of backbone network weights, only using Driver RPN generates candidate suggestion boxes ;

[0138] The two sets of features are aligned to a unified dimension using the RoI Align layer to obtain the foreground proposal box features. and This ensures the consistency of features between subsequent domain adversarial learning and course learning.

[0139] Furthermore, the adversarial domain generation module is used to construct cross-modal domain adversarial learning mechanisms:

[0140] Construct a cross-domain semantic attribute set and generate basic text features using the CLIP text encoder;

[0141] Computing source domain text embedding and text domain migration direction ;

[0142] The text domain transfer is converted into a visual domain feature offset using a cross-modal mapping function;

[0143] Generating pseudo-domain features and to combat losses Adversarial optimization is performed by minimizing the generator and maximizing the discriminator to accurately identify high-challenge generalization blind spots where the model is not adapted.

[0144] Furthermore, the evolutionary curriculum learning module is used to generate continuous pseudo-domain spaces in stages, from easy to difficult:

[0145] Phase 1: Setting a learnable offset vector Simple pseudo-domains are generated simply by linear combination of basic attributes;

[0146] Phase Two: Activation Expand the feature space to generate highly complex and challenging pseudo-domains;

[0147] By generating continuous pseudo-domains, we can achieve an evolutionary learning process from basic pseudo-domains to complex pseudo-domains, gradually improving the model's cross-domain generalization ability.

[0148] Furthermore, the alternating optimization training module is used to execute a two-step alternating optimization mechanism:

[0149] Step 1: Freeze the detector and optimize the pseudo-domain generator and cross-modal mapping function to complete the pseudo-domain generation optimization using adversarial loss + regularization loss;

[0150] Step 2: Freeze the pseudo-domain generator to detect loss and optimize the detector;

[0151] Meanwhile, learnable domain offset vectors are injected into the intermediate layers of Res3 and Res4 with a fixed probability of 0.5, introducing domain perturbation without destroying the feature structure of the source domain, thereby enhancing the model's domain adaptability.

[0152] Furthermore, the phased scheduling and buffer management module is used to perform two-stage training scheduling of the initial task and the offset training task:

[0153] Initial stage: Train the base detector based on the initial categories in the source domain;

[0154] Offset training phase: Perform pseudo-domain augmentation, adversarial feature mining, template update and weight fusion round by round;

[0155] The pseudo-domain sample buffer is dynamically maintained. After each round of offset training, redundant samples are removed and representative pseudo-domain samples are added to maintain a balanced distribution of the buffer and continuously improve the generalization ability.

[0156] Furthermore, the cross-domain detection inference module is used to perform end-to-end inference on images from unknown or unseen domains:

[0157] Input the image from the unknown domain into the trained system model, and directly output the object detection result;

[0158] The output includes: target category label, target bounding box coordinates, and detection confidence score, achieving target detection with single-domain training and global generalization.

[0159] This embodiment can automatically construct a continuous pseudo-domain space under single source domain training, accurately uncover generalization blind spots, and achieve stable and efficient cross-domain target detection. The system has a clear structure and a simple training process, which can effectively improve the model's adaptability to unknown domains, ensure detection accuracy and robustness, and has good engineering practicality and versatility.

[0160] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any changes, modifications, substitutions, integrations, and parameter changes made to these embodiments within the spirit and principles of the present invention, without departing from the principles and spirit of the present invention, through conventional substitutions or to achieve the same function, fall within the scope of protection of the present invention.

Claims

1. A single-domain generalization object detection method based on evolutionary cues against, characterized in that, include: The source image is subjected to basic enhancement processing, including random cropping, Gaussian noise, color dithering, and normalization, to obtain the enhanced image. The original image of the source domain and the enhanced image are input into a CLIP-ResNet-101 detection network based on a Faster R-CNN framework sharing weights, and visual features containing medium-high layer semantic information are extracted with ; Utilizing The driving area proposal network generates a candidate proposal frame , based on the candidate proposal frame , the candidate region features are extracted respectively from and by the RoI Align layer and ; The foreground suggestion box features are input into the adversarial domain generation module. Text features are constructed based on the cross-domain semantic attribute pool and CLIP text encoder. The text domain migration direction is calculated and mapped to visual feature offset. Pseudo-domain features are generated and pseudo-domain features with high detection loss are selected. The pseudo-domain features are input into the evolutionary course learning module, and the linear combination of basic attributes and the expansion of offset vectors are performed in stages to generate a continuous pseudo-domain space from easy to difficult. An alternating optimization mechanism is adopted. In the pseudo-domain generation stage, the pseudo-domain cue parameters and cross-modal offset mapping function are updated based on adversarial loss, domain offset constraint loss and regularization loss. In the detection training stage, the target detection network parameters are updated based on detection loss. Learnable offsets are injected into Res3 and Res4 layers with fixed probability. Multi-step training is performed in stages according to the initial task and offset, and the pseudo-domain sample buffer is dynamically updated. Input the image of the unknown domain into the trained model, and output the target category label, bounding box coordinates, and confidence score.

2. The single-domain generalization object detection method based on evolutionary hint contrastive according to claim 1, wherein, The basic enhancement process includes: The source domain image is sequentially subjected to fixed-scale random cropping, Gaussian noise addition with preset intensity, color jitter within a fixed range, and mean-variance normalization. Enhancement operations only change the statistical distribution of the image appearance, without altering the target size, location, or semantic information of its category; The enhanced image maintains the same resolution and channel dimension as the original image, and together they form the input for dual-path feature extraction.

3. The single-domain generalization object detection method based on evolutionary hint-based adversarial according to claim 2, characterized in that, The dual-path feature extraction and proposal box alignment includes: The pre-trained CLIP visual encoder is used to extract features of the source domain image and the enhanced image in parallel to obtain With , is the visual feature map of the source domain original image, is the multi-scale visual feature map of the source domain image after appearance enhancement processing; Only to Input RPN generation suggestion box ; Will , With input RoI Align layer, get the unified dimension foreground proposal box features With .

4. The single-domain generalization object detection method based on evolutionary hint contrastive according to claim 1, characterized in that, The adversarial domain generation module includes: Building a cross-domain semantic attribute set wherein, represents a semantic attribute for describing a cross-domain appearance change; inputting the cross-domain semantic attribute set into a frozen CLIP text encoder , obtaining a basic text feature through the CLIP text encoder ; based on the base text feature set , by the pseudo-domain generator F generating pseudo-domain text features : ; According to source domain text embedding = ( ) with the pseudo domain text feature Computing normalized text domain migration direction: = ; and construct ideal pseudo-domain features , the formula is: , wherein is a source domain visual feature, is a domain migration strength; by a cross-modal offset mapping function migrating direction of the text domain visual offset of the ith visual feature layer: = ( ); and inject the visual shift into the source domain visual features , to obtain actual pseudo-domain visual features: = + ; The actual pseudo-domain embedding is obtained through a detection backbone network based on the actual pseudo-domain visual features. ; Constructing adversarial losses: The minimax adversarial optimization strategy is adopted to make the pseudo-domain generator F Search for challenging pseudo-domain text features and apply the cross-modal offset mapping function. The learning process converts the text domain migration direction into the corresponding visual feature offset.

5. The single-domain generalized target detection method based on evolutionary cueing adversarial methods according to claim 4, characterized in that, The cross-modal mapping and visual offset injection include: Direction of text field migration Input cross-modal mapping function Output a domain offset vector with the same dimension as the visual features; The domain offset vectors are superimposed channel by channel onto the output feature maps of ResNet's Res3 and Res4 layers; Offset injection is randomly enabled with a probability of 0.5, introducing domain perturbation without destroying the source domain feature structure, thereby improving the model's adaptability to domain offset.

6. The single-domain generalized target detection method based on evolutionary cueing adversarial methods according to claim 1, characterized in that, The pseudo-domain generation method of the evolutionary curriculum learning module is as follows: ,in Based on basic text features, For attribute weights, A learnable offset vector; Phase 1 Only optimize The pseudo-domain is generated by a linear combination of basic attributes; Phase Two Activation This expands the feature space and generates complex pseudo-domains, enabling learning from easy to difficult courses.

7. The single-domain generalized target detection method based on evolutionary cueing adversarial methods according to claim 1, characterized in that, The alternating optimization includes two steps: Step 1: Pseudo-domain generation optimization ,in To generate loss for adversarial domains, For pseudo-domain diversity regularization loss; Step 2: Detector optimization ,in For source domain feature maps, It is a pseudo-domain transformation function. For labeling.

8. The single-domain generalized target detection method based on evolutionary cueing adversarial methods according to claim 1, characterized in that, Overall model optimization includes: The detection classification loss, bounding box regression loss, adversarial pseudo-domain generation loss, and pseudo-domain regularization loss are weighted and fused to obtain the overall optimization objective. During the detector optimization phase, the target detection network parameters are updated based on the detection classification loss and bounding box regression loss. During the pseudo-domain generation optimization stage, the pseudo-domain generator, cross-modal offset mapping function, and learnable visual offset parameters are updated based on the adversarial pseudo-domain generation loss and regularization loss. By alternately executing the detector optimization phase and the pseudo-domain generation optimization phase, the target detection network gradually adapts to the continuously generated pseudo-domain features, which range from easy to difficult.

9. The single-domain generalized target detection method based on evolutionary cueing adversarial methods according to claim 1, characterized in that, The phased training and offset update include: In the first stage, the learnable residual perturbation vector is frozen or set to zero. Based solely on the basic text feature set Optimize attribute combination weights Generate initial pseudo-domain text features located in the basic attribute space; In the second stage, a learnable residual perturbation vector is enabled based on the initial pseudo-domain text features obtained in the first stage. This expands the search range of pseudo-domain text features and generates more complex and challenging pseudo-domains. For each candidate pseudo-domain, calculate its adversarial loss, and select the top K highly challenging pseudo-domains based on the loss magnitude to participate in subsequent optimization; Based on the loss gain before and after enabling the residual perturbation vector To determine whether the residual disturbance is effective, when Greater than the preset threshold If the corresponding pseudo-domain candidate is retained, its weight in training should be reduced or it should be removed. During the update process of the pseudo-domain sample buffer, orthogonal regularization constraints are used to maintain the diversity among different pseudo-domain candidates in order to stabilize the expansion process of the continuous pseudo-domain space.

10. A single-domain generalized target detection system based on evolutionary cueing adversarial mechanisms, characterized in that, include: The data preprocessing module is used to perform random cropping, Gaussian noise, color dithering and normalization enhancement processing on the source domain image to obtain the enhanced image; The dual-path feature extraction and candidate box alignment module is used to extract source domain visual features and enhanced visual features from the CLIP-ResNet-101 detection network based on the Faster R-CNN framework, which inputs the source domain original image and the enhanced image with shared weights. The module generates candidate proposal boxes based on the source domain visual features and then aligns the source domain visual features and enhanced visual features with candidate region features through the RoI Align operation. The adversarial pseudo-domain generation module is used to construct a cross-domain semantic attribute set, input the cross-domain semantic attribute set into the frozen CLIP text encoder to obtain basic text features, generate pseudo-domain text features based on the pseudo-domain generator, calculate the text domain migration direction, and map the text domain migration direction to visual feature offset through the cross-modal offset mapping function to generate pseudo-domain visual features and perform adversarial optimization. The evolutionary curriculum learning module is used to generate pseudo-domain text features in stages. In the first stage, a basic pseudo-domain is generated based on a linear combination of basic text features. In the second stage, a learnable residual perturbation vector is introduced to expand the pseudo-domain search space, and the difficulty of the pseudo-domain is dynamically adjusted according to the loss gain and orthogonal regularization constraints. An alternating optimization training module is used to alternately perform pseudo-domain generation optimization and detector optimization. In the pseudo-domain generation optimization stage, the pseudo-domain generator, cross-modal offset mapping function and learnable visual offset parameters are updated. In the detector optimization stage, the object detection network parameters are updated, and learnable visual offsets are injected into at least one mid-to-high-level visual feature layer. The cross-domain detection inference module is used to input images from unknown domains into the trained target detection network, perform forward inference, and output the target category, bounding box coordinates, and confidence score.

Images