Efficient single-stage domain adaptive target detection method for open scene

Abstract

The invention discloses an efficient single-stage domain adaptive target detection method for an open scene. The method comprises the following steps: acquiring a to-be-detected open scene target domain image; inputting the target domain image into a single-stage domain adaptive model, wherein the single-stage domain adaptive model is a model obtained by circularly executing cognitive closed-loop iterative training including a learning step, a practice step and a reflection step on the label-free target domain data; and performing target detection processing on the target domain image based on the single-stage domain adaptive model to obtain bounding box position and category information of each target in the target domain image. By constructing a single-stage cognitive closed-loop training architecture fusing the steps of learning, practice and reflection, self-iterative evolution is realized in the same detection model, the training efficiency is improved by one order of magnitude, error accumulation is effectively inhibited, and the detection accuracy is improved. The problems of low training efficiency, bloated architecture and performance limitation caused by lack of a feedback correction mechanism due to adoption of a complex multi-stage teacher-student framework in an existing domain self-adaptive method are solved.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and in particular to an efficient single-stage domain adaptive target detection method for open scenes. Background Technology

[0002] Object detection, as a core task in this field, has wide applications in real-world open scenarios such as autonomous driving, drone inspection, and robot navigation. With the development of deep learning technology, models such as YOLO and Faster R-CNN have demonstrated excellent performance in environments consistent with their training data distribution. However, when these models are deployed to new target domains (e.g., different weather conditions, lighting environments, or geographical scenes), the significant domain shift between the source and target domains often leads to a catastrophic drop in model performance, severely hindering the practical application of object detection technology in real-world open scenarios.

[0003] To address the aforementioned domain offset problem, domain-adaptive object detection techniques have emerged. Current mainstream methods generally employ a multi-stage teacher-student framework, typically including two main stages: pre-training and self-training. This paradigm first trains an initial teacher model on the labeled source domain; subsequently, in the self-training stage, the teacher model generates pseudo-labels for the unlabeled target domain data, using these pseudo-labels as supervisory signals to iteratively train the student model. Since the initial pseudo-labels inevitably contain noise, the core of the subsequent process lies in the selection and optimization of pseudo-labels. Existing methods mainly employ confidence-based thresholding strategies (such as fixed or dynamic thresholds), while some studies have also introduced uncertainty measures (such as quantifying model uncertainty through Monte Carlo Dropout) or combined strategies such as adversarial training and domain transformation to improve model adaptability. In recent years, with breakthroughs in large-scale vision-language models (such as CLIP and Grounding DINO), advanced methods such as COIN have emerged. These methods further improve cross-domain performance by integrating the open-vocabulary detection and zero-shot classification capabilities of such models to generate and optimize pseudo-labels. Nevertheless, methods such as COIN still follow the traditional multi-stage teacher-student paradigm and introduce an additional consistency correction network, resulting in a still complex overall architecture.

[0004] Despite some progress in existing methods, they are still limited by the inherent defects of the multi-stage paradigm, mainly in the following aspects: First, the multi-stage teacher-student framework requires training and maintaining multiple models separately, resulting in a bloated overall architecture, huge computational overhead, and long training cycles, making it difficult to meet the stringent requirements of rapid deployment and real-time performance in real-world scenarios; Second, in the existing process, data and noise are transmitted unidirectionally, lacking an effective closed-loop feedback mechanism. Pseudo-label noise generated in the preceding stages (such as category confusion and positioning bias) is directly transmitted and accumulated in subsequent stages, limiting the model's performance ceiling and self-evolution ability; Finally, pseudo-label selection strategies are generally simplistic, mainly relying on local confidence thresholds while ignoring the global uncertainty (information entropy) of the predicted probability distribution. Therefore, they cannot effectively identify and filter out ambiguous samples that are "high-scoring but highly uncertain," resulting in a large number of low-quality pseudo-labels being retained, which in turn contaminates the training data and misleads the model optimization direction.

[0005] Therefore, in response to the urgent need for efficient, robust, and rapidly adaptive detection systems in real-world open scenarios, existing domain-adaptive target detection methods still have significant shortcomings in terms of architectural efficiency, error control, and noise suppression. There is an urgent need for a new domain-adaptive target detection framework that is more efficient, accurate, and capable of continuous self-evolution. Summary of the Invention

[0006] The purpose of this invention is to provide an efficient single-stage domain adaptive object detection method for open scenarios. By constructing a single-stage cognitive closed-loop training architecture that integrates "learning-practice-reflection" steps, it achieves self-iterative evolution within the same detection model, thereby improving training efficiency by an order of magnitude and effectively suppressing error accumulation. This solves the problems of low training efficiency, bloated architecture, and performance limitations caused by the lack of feedback correction mechanisms in existing domain adaptive methods due to the use of complex multi-stage teacher-student frameworks.

[0007] To address the aforementioned technical problems, a first aspect of this invention provides an efficient single-stage domain adaptive target detection method for open scenarios, comprising the following steps: Obtain the target domain image of the open scene to be detected; The target domain image is input into a single-stage domain adaptive model, which is a model obtained by iteratively training a cognitive closed loop that includes learning steps, practice steps and reflection steps on unlabeled target domain data. The target domain image is processed by the single-stage domain adaptive model to obtain the bounding box position and category information of each target in the target domain image.

[0008] Furthermore, before inputting the target domain image into the single-stage domain adaptive model, the method further includes: Retrieve unlabeled target domain data; Build or obtain the initial target detector model; Based on the target domain data, a single-stage cognitive closed-loop iterative process is executed cyclically to train and update the target detector model. The single-stage cognitive closed-loop iterative process includes learning steps, practice steps, and reflection steps in sequence. The iteration stops when the target detector model meets the convergence condition, thus obtaining the single-stage domain adaptive model.

[0009] Furthermore, the iterative process of repeatedly executing a single-stage cognitive closed-loop process to train and update the target detector model includes: In the first iteration, a pseudo-label set is generated based on the target domain data and the prediction-trained visual-language large model with open vocabulary detection capability. The pseudo-label set contains the predicted target bounding box coordinates and their corresponding category probability distributions. In the current iteration, the currently held pseudo-label set is used as a supervision signal to supervise the training of the target detector model, thus completing the learning step. Using the target detector model updated through the learning steps, forward reasoning is performed on the unlabeled target domain data to generate the detection result set for the current round, thus completing the practical steps. The detection result set of the current round is compared and verified with the pseudo-label set used as input to the learning step of the current round, and the detection results are corrected according to the verification results to generate an updated pseudo-label set, thus completing the reflection step. The updated pseudo-label set is used as the input for the next iteration, replacing the currently held pseudo-label set, and the learning step, the practice step, and the reflection step are repeated.

[0010] Furthermore, before using the currently held pseudo-label set as a supervision signal to perform supervised training on the target detector model, the process further includes: For each pseudo-label in the currently held pseudo-label set, based on the category probability distribution it contains, calculate the local confidence level reflecting the category certainty and the global information entropy reflecting the degree of dispersion of the probability distribution; By fusing the local confidence score with the global information entropy, the uncertainty perception score of the pseudo-label is calculated; The uncertainty perception score is compared with a preset quality threshold, and pseudo-labels with scores better than the quality threshold are retained to form a high-quality pseudo-label set for supervised training.

[0011] Furthermore, the uncertainty-perceived score of the pseudo-label. The calculation formula is: ; ; ; in, The global information entropy, The local confidence level is... Let be the probability value of the i-th category of the pseudo-label, and s be the preset number of categories in the target domain. For uncertain adjustment term weights, This is the sensitivity scaling factor. The slope parameter of the Sigmoid function. The weight of the confidence penalty term. This is the confidence decay index.

[0012] Further, the step of comparing and verifying the detection result set of the current round with the pseudo-label set used as input to the learning step of the current round, and correcting the detection results based on the verification results, includes: The detection result set of the current round is matched with the pseudo-label set for spatial and category consistency. Based on the overlap of bounding boxes and the consistency of category labels, the matching results are divided into a consistent detection subset and an inconsistent detection subset. For the inconsistency detection subset, a preset cross-modal semantic alignment model is invoked to calculate the semantic similarity between the target region image and the candidate category text description in the inconsistency detection subset, and the conflicting category labels are corrected based on the semantic similarity. Based on the preset prior logic rules that reflect the co-occurrence or mutual exclusion relationship between target scenes, the tag set after being corrected by the preset cross-modal semantic alignment model is verified to obtain the updated pseudo-tag set.

[0013] Further, the step of performing spatial and category consistency matching between the detection result set of the current round and the pseudo-label set includes: Calculate the intersection-union ratio (IUU) between the bounding box of each detection result in the current round of the detection result set and the bounding box of each pseudo-label in the pseudo-label set; The detection results with the same category label and the intersection-union ratio greater than the preset overlap threshold are divided into the consistent detection subset along with the pseudo-label pairs; Detection results or pseudo-labels that have an intersection-union ratio greater than the preset overlap threshold but have different category labels, or that fail to reach the preset overlap threshold with any element in the other set, are classified into the inconsistency detection subset.

[0014] Further, calculating the semantic similarity between the target region image and the candidate category text description in the inconsistency detection subset includes: The image encoder of the preset cross-modal semantic alignment model extracts the image embedding vector of the target region image; The text encoder of the preset cross-modal semantic alignment model extracts the text embedding vector of the candidate category text description; The cosine similarity between the image embedding vector and the text embedding vector is calculated as the semantic similarity.

[0015] Furthermore, the correction of conflicting category labels based on the semantic similarity includes: For the detection results and pseudo-label pairs with category conflicts in the inconsistency detection subset, calculate the semantic similarity between the corresponding target region image and the text description corresponding to the conflict category; If the category with the highest semantic similarity is a preset category among the conflicting categories, then the corresponding label is corrected to the preset category; otherwise, the detection results and pseudo-label pairs with category conflicts are discarded.

[0016] Furthermore, the verification of the tag set after correction by the preset cross-modal semantic alignment model, based on the preset prior logic rules reflecting the co-occurrence or mutual exclusion relationship between targets, includes: If a target category of pedestrian is detected, and there are bicycle or motorcycle targets in the vicinity of the pedestrian, the pedestrian target category is changed to cyclist. If a target category of cyclist is detected, and there are no bicycle or motorcycle category targets in its spatial neighborhood, the cyclist target category is corrected to pedestrian. If both pedestrian and cyclist category labels exist for the same spatial location, the cyclist category label will be retained and the pedestrian category label will be deleted.

[0017] Accordingly, a second aspect of the present invention provides an electronic device, including: at least one processor; and a memory connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to cause the at least one processor to perform the above-described efficient single-stage domain adaptive target detection method for open scenes.

[0018] Accordingly, a third aspect of the present invention provides a computer-readable storage medium having computer instructions stored thereon, which, when executed by a processor, implement the above-described efficient single-stage domain adaptive target detection method for open scenarios.

[0019] The above-described technical solutions of the embodiments of the present invention have the following beneficial technical effects: 1. By abandoning the traditional multi-stage teacher-student separation training paradigm, we innovatively propose and implement a single-stage self-evolutionary closed-loop training architecture inspired by human cognition. This architecture drives the same lightweight detection model to cyclically perform cognitive iterations of "learning, practice, and reflection" on the target domain data, realizing the model's autonomous and continuous optimization. It greatly simplifies the training process, eliminates the path of error propagation and accumulation between stages, thereby improving the efficiency of domain adaptive training by an order of magnitude while maintaining high detection accuracy, and significantly reducing model complexity and computational cost, making it possible to quickly deploy and apply the technology on resource-constrained edge devices. 2. To address the initial pseudo-label noise problem, an uncertainty-aware gating screening mechanism is introduced in the learning step; a breakthrough is made by fusing local confidence and global information entropy for joint evaluation, constructing a refined uncertainty-aware scoring function that can effectively identify and eliminate ambiguous or deceptive samples with "high scores but high uncertainty" that traditional methods cannot distinguish; the quality of pseudo-labels used for model supervised training is significantly improved from the source, providing a more reliable knowledge base for the subsequent iterative evolution of the model, thereby directly improving the initial performance of the model in the target domain and the stability of subsequent optimization; 3. In the reflection step, a multi-level, multi-source knowledge closed-loop correction system is constructed. Cognitive conflicts are identified through consistency matching, and powerful external visual-language models such as CLIP are introduced for zero-shot arbitration. Heuristic rules based on scenario priors are used for logical verification, forming a complete self-correction process of "discovering conflicts - adjudicating and correcting - refining logic". This effectively simulates the human ability to reflect and correct, enabling the model to continuously correct historical cognitive biases using newly acquired knowledge. This not only significantly improves the model's detection accuracy and robustness in various complex cross-domain scenarios, but also endows it with the core ability to continuously self-evolve and adapt to open and unknown environments. Attached Figure Description

[0020] Figure 1 This is a flowchart of an efficient single-stage domain adaptive target detection method for open scenarios provided by an embodiment of the present invention; Figure 2 This is a schematic diagram of the overall algorithm framework provided in the embodiments of the present invention; Figure 3 This is a schematic diagram of the training process of a single-stage domain adaptive model provided in an embodiment of the present invention; Figure 4 This is a flowchart of the uncertainty perception gating mechanism provided in the embodiments of the present invention; Figure 5 This is a comparison chart of detection results provided in an embodiment of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments and the accompanying drawings. It should be understood that these descriptions are merely exemplary and not intended to limit the scope of the invention. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concept of the invention.

[0022] Please refer to Figure 1 , Figure 2 and Figure 3 The first aspect of this invention provides an efficient single-stage domain adaptive target detection method for open scenarios, comprising the following steps: Step S100: Obtain the target domain image of the open scene to be detected.

[0023] In one embodiment of this invention, the open-scene target domain image refers to a real-world environment image that differs significantly in distribution from the initial training data of the model. Examples include road scenes captured by autonomous vehicles under adverse weather conditions such as fog or rain, or images taken by drones during inspections in unknown geographical areas, showing different lighting, backgrounds, and target appearances. These images collectively constitute the target domain that the model needs to adapt to and complete the detection task. This step is the starting point for subsequent adaptive detection processes; the images acquired will serve as direct input to the single-stage domain adaptive model, used to evaluate and verify the model's actual performance and generalization ability after deviating from the original training distribution.

[0024] Step S300: Input the target domain image into the single-stage domain adaptive model. The single-stage domain adaptive model is a model obtained by iteratively training a cognitive closed loop that includes learning steps, practice steps, and reflection steps on unlabeled target domain data.

[0025] The aforementioned single-stage domain adaptation model is a unified object detection network (such as a lightweight YOLO series network) specifically designed for achieving efficient domain adaptation in open scenarios. It is not a static, pre-trained general model, but a state that has been specifically adapted to the current target domain environment through a dynamic self-evolution process.

[0026] The process of obtaining this model, i.e., its training method, constitutes the core innovation of this approach. Specifically, the single-stage domain adaptive model is ultimately converged by iteratively executing a cognitive closed-loop optimization process on a target domain dataset without any manual annotations. This process is tailored to specific application scenarios (e.g., adapting the model from a clear city street view data domain to a foggy driving scenario target domain), and then enters a closed-loop iteration: In the learning step, a high-quality pseudo-label is selected through an uncertainty-aware gating mechanism to supervise the training of the single-stage domain adaptive model, enabling it to initially learn the features of foggy targets; in the practice step, the updated model is used to re-infer all foggy data, generating a new round of detection results; in the reflection step, the new results are matched with historical pseudo-labels for consistency, and semantic knowledge provided by external models such as CLIP is used to adjudicate and correct inconsistencies and conflicts in detections, thereby generating a higher-quality pseudo-label set. This "learning-practice-reflection" closed loop is repeated multiple times. Throughout this process, only the parameters of this single-stage domain adaptive model are maintained and updated. With each loop closure, the model is trained using self-generated, purified, and corrected feedback signals, gradually adjusting its internal representation to better fit the data distribution (target domain) of the foggy scene until performance convergence. Therefore, the model used in step S300 is the final product, refined after this series of loop iterations, possessing strong adaptive capabilities for a specific target domain (such as driving in fog). It abandons the complex paradigm of simultaneously training and coordinating multiple independent models (teacher and student) in traditional methods, achieving a unified and efficient training process.

[0027] Step S500: Target detection processing is performed on the target domain image based on a single-stage domain adaptive model to obtain the bounding box position and category information of each target in the target domain image.

[0028] Utilizing the aforementioned single-stage domain adaptive model, which has been trained through the cognitive closed-loop iteration and thus optimized to adapt to the target domain data distribution, a forward propagation computation is performed. First, the input image is standardized and formatted. Then, the image data flows through the model's multi-layer convolutional or Transformer encoder structure for deep feature extraction, generating multi-scale feature maps that can characterize visual patterns and semantic information in the image. Based on this, at each preset anchor point or query position of the feature map, the model performs two parallel regression and classification computations: On the one hand, the bounding box regression branch predicts the precise geometric offset of the target that may exist at the corresponding position. These offsets, combined with preset anchor box parameters, are decoded and converted into the bounding box coordinates of each candidate target in the image coordinate system, i.e., its center point position, width, and height; on the other hand, the classification branch generates a class probability distribution vector for each predicted bounding box. This vector represents the probability that the target within the box belongs to each category in a predefined category set (such as "car", "pedestrian", "cyclist", etc.). Finally, the model applies a non-maximum suppression post-processing algorithm to filter all initial predictions, removing redundant boxes with excessive overlap and retaining the predictions with the highest confidence, thus outputting a structured list of detection results. Each item in this list clearly corresponds to a recognized target instance in the image, including its refined final bounding box pixel coordinates and its most likely class label. This process is efficient and direct, representing the model's application of the feature representations and discriminative abilities learned during adaptive training and optimized for the target domain to inference on real-world new samples.

[0029] In a specific embodiment of the present invention, before inputting the target domain image into the single-stage domain adaptive model in step S300, the method further includes: Step S210: Obtain unlabeled target domain data.

[0030] Unlabeled target domain data refers to a collection of real-world scene images that the model is expected to adapt to, but which lack human annotation. In a specific embodiment, if the goal is to adapt a detection model pre-trained on a sunny city street view (source domain) to a foggy driving scenario, this step requires collecting a large number of road traffic images taken under foggy conditions as the target domain dataset. These images only contain raw pixel information and do not have any bounding box coordinates or category labels for targets such as vehicles, pedestrians, and traffic signs. Their data distribution differs significantly from the source domain data, simulating the domain offset challenge faced by the model in real-world deployment.

[0031] Step S220: Build or obtain the initial target detector model.

[0032] The initial object detector model is a neural network with basic object detection capabilities, but its parameters have not yet been optimized for the target domain. In this embodiment, it can be a lightweight, single-stage detection model (such as YOLOv8s) pre-trained on a large, general dataset (such as COCO), or it can be a model with weights randomly initialized according to the selected architecture. This model serves as the sole carrier for subsequent closed-loop iterative optimization; its architecture will remain unchanged, while its parameters will be continuously updated during the iteration process to incorporate target domain characteristics.

[0033] Step S230: Based on the target domain data, a single-stage cognitive closed-loop iterative process is executed repeatedly to train and update the target detector model. The single-stage cognitive closed-loop iterative process includes a learning step, a practice step, and a reflection step in sequence.

[0034] In the single-stage cognitive closed-loop iterative process, each round of the "learning step" begins with the following: In the first iteration, a pseudo-label set is generated based on the target domain data and a large-scale visual-language model with open-vocabulary detection capabilities. This pseudo-label set undergoes quality screening based on uncertainty. The high-quality pseudo-labels are then used to supervisedly train the current target detector model, updating its parameters. In the subsequent "practice step," the updated model is used to perform forward inference on all foggy images in the target domain, generating a new set of detection results reflecting the model's current cognitive level. In the "reflection step," this new set of results is compared with the old pseudo-label set input from the previous "learning step." Matching is performed based on spatial overlap and category consistency to identify stable and consistent detections as well as conflicting ones. For conflicting detections, an external pre-trained model (such as CLIP) is introduced to calculate the semantic similarity between the image region and the category text for arbitration. Finally, scene-knowledge-based logical rules may be applied for verification, resulting in a higher-quality, updated pseudo-label set. This updated pseudo-label set serves as the input for the next iteration's "learning step," and this process is repeated continuously.

[0035] Step S240: Stop iterating when the target detector model meets the convergence condition, and obtain a single-stage domain adaptive model.

[0036] In this embodiment, the "convergence condition" can be a preset fixed number of iterations (e.g., 5 rounds), or it can be that the model's performance metrics (e.g., mean average precision, mAP) on the target domain validation set no longer show significant improvement. When this condition is met, the iterative process stops. At this point, the target detector model, after multiple rounds of "learning-practice-reflection" closed-loop optimization, has had its parameters sufficiently adjusted, and its internal feature representation can better capture and understand the characteristics of targets in foggy scenes. This model is the final "single-stage domain adaptive model" that has adapted to the target domain, possessing the ability to perform high-precision and high-efficiency target detection in corresponding foggy open scenes.

[0037] By leveraging the initial knowledge provided by a large vision-language model and designing a closed-loop iterative process that includes pseudo-label quality screening, model inference verification, and multi-source feedback correction, a single detection model can update its parameters using self-generated and continuously optimized supervision signals, ultimately converging to a state highly adapted to the target domain data distribution. This process avoids the overhead of maintaining a complex multi-stage model architecture, effectively suppresses the propagation and accumulation of noisy pseudo-labels, and achieves efficient and accurate domain-adaptive model training.

[0038] Specifically, step S230 involves repeatedly executing a single-stage cognitive closed-loop iterative process to train and update the target detector model, including: Step S241: In the first iteration, a pseudo-label set is generated based on the target domain data and the prediction-trained visual-language large model with open vocabulary detection capability. In the current iteration, the currently held pseudo-label set is used as a supervision signal to supervise the training of the target detector model, thus completing the learning step.

[0039] In the first iteration, the foggy images obtained in step S210 are batch-input into a pre-trained visual-language model (e.g., Grounding DINO) with open vocabulary detection capabilities. Based on prior knowledge learned from massive image-text pairs, this model analyzes each foggy image and outputs a series of preliminary detection results. Each result contains a predicted target bounding box (representing its position in the image in coordinates) and a class probability distribution vector (indicating the probability that the target within the box belongs to each class, such as "car" or "person"). This automatically generated, unfiltered set of (bounding box, probability distribution) pairs constitutes a pseudo-label set.

[0040] In subsequent iterations, the currently held set of pseudo-labels is used as training samples, where each sample contains the coordinates of a target bounding box and its class probability distribution. Before training begins, a quality screening step (e.g., an uncertainty-aware scoring mechanism) is applied to these pseudo-labels to filter out high-noise samples with excessively low confidence or overly flat predicted probability distributions, thus constructing a relatively clean, high-quality supervision signal set. Subsequently, the target detector model (e.g., a YOLO-based network) processes the target region features corresponding to these pseudo-labels via forward propagation and calculates loss functions (e.g., localization loss and classification loss) between its predicted output and the bounding box coordinates and class distribution specified in the pseudo-labels. Through backpropagation, this loss is used to adjust all weight parameters of the model, enabling the model to gradually learn to more accurately locate targets and predict their classes in target domain data (e.g., images of driving in foggy weather).

[0041] Step S242: Using the target detector model updated by the learning step, perform forward inference on the unlabeled target domain data to generate the detection result set for the current round, thus completing the practical step.

[0042] At this stage, the target detector model, after parameter updates in step S241, is used to perform a complete forward inference on the entire unlabeled target domain dataset (e.g., all foggy scene images). The model processes each input image, extracts features through its internal backbone network, and outputs a series of raw detection predictions from the detection head. These predictions include the bounding box coordinates, class confidence, and class probability for each candidate target. Post-processing steps such as non-maximum suppression are typically performed to filter out highly overlapping low-confidence boxes, ultimately forming a structured set of detection results for the current round that represents the model's cognitive achievements.

[0043] Step S243: Compare and verify the detection result set of the current round with the pseudo-label set used as input to the learning step of the current round, and correct the detection results according to the verification results to generate an updated pseudo-label set, thus completing the reflection step.

[0044] First, consistency matching is performed: the current round of detection results generated in step S242 is systematically compared with the old pseudo-label set input to the learning step (step S241) at the beginning of this iteration. The comparison is mainly based on spatial location (calculating the intersection-union ratio of bounding boxes) and category labels, thereby dividing the detection results into "consistent sets" (high positional overlap and the same category) and "inconsistent sets" (including location matching but category conflict, as well as isolated detections that failed to match). For inconsistent sets, an external, pre-trained cross-modal model (e.g., CLIP) is introduced to calculate the semantic similarity between the image embedding of the conflicting target region and the text embedding of the candidate category, and to adjudicate the conflicting category based on the similarity level, deciding whether to retain, modify, or discard the detection.

[0045] Furthermore, heuristic rules based on scenario priors can be applied (e.g., if a "pedestrian" target is near a "bicycle," then its category is corrected to "cyclist") for final logical verification. After the above multi-step verification and correction, a set of updated pseudo-labels with significantly improved quality is generated by combining the high-confidence results in the consistency set and the corrected inconsistent results.

[0046] In step S244, the updated pseudo-label set is used as the input for the next iteration, replacing the currently held pseudo-label set, and the learning step, practice step, and reflection step are repeated.

[0047] After reflecting and correcting in step S243, the newly generated, higher-quality updated pseudo-label set is set as the current pseudo-label set used in step S241 of the next iteration, thus replacing the old supervision signal. Subsequently, the training process jumps back to step S241, starting with the new pseudo-label set, and repeats the complete cognitive loop of "learning-practice-reflection." This cycle repeats continuously, with each iteration training the model based on the results of the previous optimization (better pseudo-labels). The better-trained model then infers and generates higher-quality detection results for the next round of correction, forming a positive feedback loop that drives continuous self-improvement of model performance.

[0048] By iteratively executing a cycle of "supervised training - inference verification - validation correction," this invention enables a single-target detector model to utilize self-generated and continuously optimized pseudo-labels as supervisory signals, achieving adaptive parameter updates. The closed-loop design ensures that the quality of the supervisory signal continuously improves during iteration, effectively suppressing the propagation and accumulation of initial noise, while gradually adjusting the model parameters to a state highly matched with the target domain data distribution.

[0049] Furthermore, please refer to Figure 4 Before using the currently held pseudo-label set as a supervision signal to perform supervised training on the object detector model in step S241, the following steps are also included: Step S2401: For each pseudo-label in the currently held pseudo-label set, calculate the local confidence level reflecting the certainty of the category and the global information entropy reflecting the dispersion of the probability distribution based on the category probability distribution it contains.

[0050] Specifically, for each pseudo-label in the current pseudo-label set, its category probability distribution vector is the output of a large visual-language model (such as Grounding DINO) making multi-class predictions for a certain region in the target domain image. To evaluate the quality of this prediction, two core metrics need to be calculated: first, local confidence, which is the maximum value in the probability distribution vector, directly reflecting the model's confidence in the most likely category determined by the pseudo-label; and second, global information entropy, calculated based on the entire probability distribution vector. The larger the value, the more uniform the probability distribution, the more hesitant the model is among multiple categories, and the higher the uncertainty. For example, in a pseudo-label for a foggy scene, if its probability distribution is [car: 0.9, truck: 0.1], then the local confidence is high (0.9) and the entropy value is low, indicating a clear prediction; if the distribution is [car: 0.4, truck: 0.3, bus: 0.3], then although the local confidence may still meet the simple threshold (0.4), its high entropy value reveals that the model has great global uncertainty in judging the target category.

[0051] Step S2402: Combine local confidence and global information entropy to calculate the uncertainty perception score of pseudo-labels.

[0052] The scoring function is based on global information entropy and dynamically scaled using a modulation factor related to local confidence. When local confidence is low, this modulation factor amplifies the penalty effect of the entropy value; conversely, when local confidence is high, the penalty effect weakens. Simultaneously, the scoring function includes a separate term that directly penalizes low confidence itself. By weighted summing the modulated entropy penalty term and the direct confidence penalty term, the uncertainty-perceived score for each pseudo-label is finally calculated. The lower the score, the higher the predictive certainty and the better the quality of the pseudo-label. This fusion mechanism ensures that the screening process considers not only which class the model deems most likely (local confidence) but also whether the model struggles to choose among multiple classes (global information entropy), thus more comprehensively identifying potential low-quality pseudo-labels.

[0053] Step S2403: Compare the uncertainty perception score with a preset quality threshold, and retain pseudo-labels with scores better than the quality threshold to form a high-quality pseudo-label set for supervised training.

[0054] A quality threshold (e.g., 0.7) is preset, which can be adjusted according to the specific situation of the target domain data. Then, the uncertainty-perceived score of each pseudo-label calculated in step S2402 is compared with this preset threshold. All pseudo-labels with scores lower than or equal to this threshold (i.e., better scores indicate lower uncertainty) are judged as high-quality, high-determinism samples and retained; while pseudo-labels with scores higher than this threshold are considered high-uncertainty, low-quality noise samples and are discarded. After this comparison and filtering operation, all retained pseudo-labels constitute a "high-quality pseudo-label set" with significantly improved purity. This set will serve as the supervision signal used in subsequent supervised training (step S241), thereby ensuring that the direction of model updates is guided by more reliable knowledge.

[0055] By jointly evaluating the local confidence and global information entropy of pseudo-labels, a refined uncertainty scoring function is constructed. This function can effectively identify and filter out deceptive noise samples that appear to meet the confidence threshold but have fuzzy actual prediction distributions and low certainty. Integrating this screening step before the learning step in the cognitive closed-loop iteration can significantly improve the purity and reliability of the supervision signals used for model training from the source.

[0056] Correspondingly, the uncertainty perception score of pseudo-labels The calculation formula is: .

[0057] .

[0058] .

[0059] in, For global information entropy, For local confidence level, Let be the probability value of the i-th category of the pseudo-label, and s be the preset number of categories in the target domain. For uncertain adjustment term weights, This is the sensitivity scaling factor. The slope parameter of the Sigmoid function. The weight of the confidence penalty term. This is the confidence decay index.

[0060] For each pseudo-label generated by the large vision-language model (using Grounding DINO in this invention), its category prediction result is defined as a probability distribution vector. (Where s is the preset number of categories in the target domain, such as s=8 in the Foggy Cityscapes dataset, containing categories such as car, person, and rider), and two core evaluation metrics are calculated based on this vector: Confidence level c: the maximum value in the probability distribution, i.e. This reflects the local classification of the pseudo-label. If c is close to 1, it indicates that the model is highly confident in predicting that category; Information entropy H(p): measures the global uncertainty of a probability distribution, and is calculated using the following formula: If H(p) is close to 0, it indicates that the probability distribution is concentrated in a certain category, and the uncertainty of the pseudo-label is low; if H(p) is large, it indicates that the category prediction is dispersed, and the uncertainty of the pseudo-label is high.

[0061] (2) Uncertainty perception score function S(p) To achieve dual-dimensional collaborative evaluation, an uncertainty-aware scoring function S(p) is constructed. The quality of pseudo-labels is quantified by a weighted fusion of the uncertainty modulation term and the confidence penalty term (the smaller S(p), the higher the quality of the pseudo-labels). The formula is as follows: , The default values ​​for each parameter are as follows: .

[0062] The formula consists of two co-operating terms. The first term is the uncertainty modulation term. In this term, the entropy H(p) is scaled by a confidence-related factor, where the expression is... The first term is a sigmoid-like function whose value decreases as the confidence level *c* increases. Taking the logarithm of this function increases its slope, thereby improving its sensitivity to changes in *c*. Therefore, this term penalizes high entropy values, and this penalty is dynamically amplified for detections with low confidence. The second term is a confidence penalty term, which directly penalizes low confidence levels. This term ensures that even low-entropy predictions are penalized if their absolute confidence is insufficient.

[0063] Further, step S243, which compares and verifies the detection result set of the current round with the pseudo-label set used as input to the learning step of the current round, and corrects the detection results based on the verification results, includes: Step S2431: Perform spatial and category consistency matching between the detection result set of the current round and the pseudo label set. Based on the overlap of the bounding boxes and the consistency of the category labels, divide the matching results into a consistent detection subset and an inconsistent detection subset.

[0064] Specifically, the set of detection results for the current round output by the practice step (step S242) is systematically compared with the set of pseudo-labels held before the start of the current learning step (step S241). The matching operation is based on two technical indicators: first, the degree of overlap in spatial location, which is usually achieved by calculating the intersection-union ratio (IU) of two bounding boxes; and second, whether the attached category labels are completely identical. In the embodiment of the foggy driving scenario, if the current model detects a bounding box whose IU with a box in the pseudo-label set exceeds a preset threshold (e.g., 0.5), and both are identified as "car", then the detection pair is considered to be consistent in both space and category and is classified into the "consistent detection subset". Conversely, if two boxes have high overlap but different category labels (e.g., one labeled "car" and the other "truck"), or if a detection box cannot find a sufficiently overlapping corresponding box in another set, then these cases are classified into the "inconsistent detection subset".

[0065] Step S2432: For the inconsistency detection subset, a preset cross-modal semantic alignment model is invoked to calculate the semantic similarity between the target region image and the candidate category text description in the inconsistency detection subset, and the conflicting category labels are corrected based on the semantic similarity.

[0066] In practice, a pre-trained cross-modal semantic alignment model, such as CLIP, is first invoked. For each item to be decided in the inconsistency subset (e.g., a spatially matched but class-conflicting "car-truck" pair, or an isolated "pedestrian" detection that is not matched), an image patch of its target region is extracted. Simultaneously, the relevant candidate categories (e.g., "car" and "truck" in the conflicting pair, or "pedestrian" in the isolated detection) are converted into text descriptions (e.g., "a photo of a car"). Subsequently, the image encoder and text encoder of CLIP are used to calculate the feature embeddings of the image patch and the feature embeddings of the text descriptions of each category, and the cosine similarity between them is calculated as the semantic similarity score. The correction decision is based on similarity comparison: for class-conflicting pairs, if the semantic similarity of one category is significantly higher than that of the other, the label is corrected to the high-similarity category; if the similarity of the two is close or neither is high, the unreliable detection pair may be discarded. For isolated detections, if the semantic similarity corresponding to their own category label is higher than a certain threshold, they are retained; otherwise, they are discarded.

[0067] Step S2433: Based on the preset prior logic rules that reflect the co-occurrence or mutual exclusion relationship between the target scenes, the tag set after being corrected by the preset cross-modal semantic alignment model is verified to obtain the updated pseudo tag set.

[0068] The system applies a set of pre-defined prior logic rules that encode common co-occurrence or mutual exclusion relationships between targets. In an embodiment of an autonomous driving scenario, typical rules include: if a target is classified as "pedestrian," and there are "bicycle" or "motorcycle" targets in its spatial neighborhood (determined by intersection-union ratio), then based on the common sense that "cyclists usually co-occur with vehicles," the category of the "pedestrian" target is corrected to "cyclist." Conversely, if no "bicycle" or "motorcycle" is detected around a "cyclist" target, it may be corrected to "pedestrian." Furthermore, if spatial overlap determines that the same target's location contains both "pedestrian" and "cyclist" labels, then only the "cyclist" label is retained based on mutual exclusion relationships. These scenario-based prior rules perform final filtering and adjustment on the output of step S2432, further eliminating labeling errors that do not conform to real-world logic, thereby generating an "updated pseudo-label set" with higher logical consistency.

[0069] By first performing internal consistency matching to lock in conflicts and uncertainties, then introducing general semantic knowledge from external cross-modal models for arbitration, and finally refining the results by integrating scenario-specific logical rules, the accuracy and logical rationality of pseudo-labels are systematically improved. This mechanism simulates the process of cross-validation and decision optimization using multi-source information, effectively correcting cognitive biases and noise generated during model adaptation, and providing significantly improved supervisory signals for the next iteration of learning.

[0070] Furthermore, step S2431, which involves performing spatial and class consistency matching between the current round's detection result set and the pseudo-label set, includes: Step S2431a: Calculate the intersection-union ratio (IUU) between the bounding box of each detection result in the current round's detection result set and the bounding box of each pseudo-label in the pseudo-label set.

[0071] For each detection result in the current round's detection result set, it contains a bounding box predicted by the model (usually represented by center point coordinates, width, and height). Similarly, each pseudo-label in the pseudo-label set also contains a bounding box generated or corrected by the previous round's process. This step requires traversing these two sets and calculating the intersection-union ratio (IU) for each possible combination (i.e., a current detection box and a historical pseudo-label box). The IU is a common metric in computer vision that measures the degree of overlap between two rectangular regions; its value is the ratio of the intersection area to the union area of ​​the two boxes. In a specific embodiment, such as a foggy driving scenario, calculating the IU between the current model's predicted bounding box for a blurred vehicle target and the corresponding labeled bounding box in the historical pseudo-labels directly reflects the consistency and stability of the model's localization of the same target and is the primary geometric basis for determining whether the two point to the same physical entity in the image.

[0072] Step S2431b: The detection results with an intersection-union ratio greater than the preset overlap threshold and the same category label are divided into a consistent detection subset along with the pseudo-label pairs.

[0073] The system presets an overlap threshold (e.g., 0.5) to determine whether two bounding boxes are spatially close enough to be considered the same target. After completing all intersection-union (IU) calculations in step S2431a, this step performs a filtering: for detection results-pseudo-label pairs with IU values ​​greater than the preset overlap threshold, the accompanying category labels (e.g., "car" or "pedestrian") are further examined. If the two labels are completely identical, it indicates that the current model's prediction is highly consistent with historical knowledge not only in target localization but also in semantic category judgment. In foggy scenarios, this means, for example, an area historically labeled as "car" with a relatively accurate location, is detected again in the current round by the model with a highly overlapping bounding box and the same "car" category. Such detection pairs are considered robust and reliable parts of the model's cognition and are therefore classified into the "consistent detection subset." This subset represents knowledge reinforced during iteration and is usually directly retained to construct the supervision signal for the next round.

[0074] Step S2431c: Detection results or pseudo-labels that have an intersection-union ratio greater than a preset overlap threshold but have different category labels, as well as those that fail to reach a preset overlap threshold with any element in the other set, are classified into an inconsistent detection subset.

[0075] The classification is also based on the comparison of intersection-union ratio (IU) and class labels, but the results do not meet the dual consistency criteria of step S2431b, specifically falling into two categories: The first category is "spatial matching but class conflict," where the IU of two bounding boxes is greater than the preset overlap threshold, but their class labels are different (e.g., the historical pseudo-label is "truck," while the current detection result is "bus"). In low-visibility scenarios such as foggy weather, this type of conflict is often caused by blurred target appearance and unclear model classification boundaries. The second category is "unmatched isolated items," where the bounding box of a certain detection result (or pseudo-label) cannot find any bounding box in the other set whose IU reaches the preset threshold. This may correspond to newly detected targets (true positives or false detections), targets missed in historical annotations (missed detections), or targets that cannot be associated due to excessive localization deviation. Both of the above categories are classified into the "inconsistent detection subset," which encompasses all uncertain information in the current cognitive loop that requires further arbitration, verification, or cleanup.

[0076] By calculating the geometric intersection-union ratio (GUC) and comparing class labels, the output of the current model is finely compared with historical supervision signals, thereby objectively distinguishing stable and consistent cognitions from conflicting or uncertain parts. This process provides clear input for the subsequent reflection and correction phase: the consistent detection subset is consolidated as highly reliable knowledge, while the inconsistent detection subset is directed to a more complex multi-source verification process. Through this structured information diversion mechanism, this invention can focus computational resources on processing truly ambiguous samples, ensuring both the effective reuse of reliable knowledge and targeted handling of noise and conflicts.

[0077] Furthermore, the semantic similarity calculation in step S2432 between the target region image and the candidate category text description in the inconsistency detection subset includes: Step S2432a: Extract the image embedding vector of the target region image through the image encoder of the preset cross-modal semantic alignment model.

[0078] For each target region to be adjudicated (e.g., a conflict region in a foggy image labeled "truck" but detected as "bus" by the current model), a corresponding image patch is first cropped from the original image based on its bounding box coordinates. This image patch is then input into a pre-defined cross-modal semantic alignment model (such as CLIP) image encoder. This encoder is typically a deep convolutional neural network or a visual Transformer network. Through a series of non-linear transformations and downsampling, it extracts multi-level visual features contained within the image patch and finally compresses and aggregates them into a fixed-dimensional, high-level semantic feature vector, i.e., the image embedding vector. In foggy scenes, this process can encode the blurred vehicle appearance, specific contours, and texture information into a compact numerical vector that contains the deep visual semantics of the image region.

[0079] Step S2432b involves extracting the text embedding vectors of candidate category text descriptions using a text encoder based on a preset cross-modal semantic alignment model.

[0080] The candidate category text descriptions that the system needs to process typically come from the category names involved in the current conflict (such as "truck" and "bus"), or the category labels of isolated detections (such as "pedestrian"). To make the text information processable by the model, these category names are first constructed as natural language cues, such as "a photo of a truck". These text descriptions are then fed into the text encoder of the same cross-modal semantic alignment model (such as CLIP). This encoder is typically a Transformer-based language model that transforms the input text sequence into a series of context-dependent features and ultimately outputs a text embedding vector with the same dimension as the image embedding vector. This vector encodes the semantic information of the category (such as attributes related to the "truck" concept) into numerical form, allowing direct similarity measurement between image content and text concepts in the same vector space.

[0081] Step S2432c: Calculate the cosine similarity between the image embedding vector and the text embedding vector as the semantic similarity.

[0082] After obtaining the image embedding vector of the target region image (step S2432a) and the text embedding vector of the candidate category text description (step S2432b), the cosine similarity between these two vectors is calculated. Cosine similarity is an indicator that measures the degree of alignment between two vectors in direction, with a value ranging from -1 to 1. The closer the value is to 1, the more consistent the semantic direction of the two vectors. Specifically, it is calculated as the dot product of the two vectors divided by the product of their respective moduli (norms). In the adjudication example of a foggy driving scenario, the similarity between the conflict region image embedding and the "truck" text embedding, and the similarity with the "bus" text embedding, are calculated respectively. These two values ​​directly and objectively reflect the semantic closeness of the blurred target region to the text concepts of "truck" and "bus," respectively, thus providing quantifiable external evidence based on large-scale pre-trained knowledge for determining its true category.

[0083] By extracting image embeddings of conflicting visual regions and text embeddings of candidate categories separately, and calculating cosine similarity within a unified semantic space, the abstract classification conflict problem is transformed into a computable vector similarity comparison problem. This process fully leverages the general visual-language association knowledge learned by cross-modal pre-trained models on massive amounts of image and text data, providing a powerful, zero-shot external adjudication basis for class ambiguity caused by appearance blurring in specific target domain scenarios (such as foggy weather). Through this objective semantic similarity metric, the system can effectively clarify and correct class conflicts in inconsistency detection, or verify the credibility of isolated detections, thereby significantly improving the semantic accuracy and reliability of the generated updated pseudo-label set.

[0084] Furthermore, step S2432, which corrects conflicting category labels based on semantic similarity, includes: Step S2432d: For the detection results and pseudo-label pairs with category conflicts in the inconsistency detection subset, calculate the semantic similarity between the corresponding target region image and the text description corresponding to the conflict category.

[0085] For each pair of detection results and pseudo-labels identified as having category conflicts in the inconsistency detection subset (e.g., in a foggy image, a historical pseudo-label identifies a region as a "truck," while the current model detects it as a "bus"), it is necessary to obtain the objective arbitration basis for the conflict pair. First, based on the bounding box coordinates of the same target region associated with the pair of detections, the corresponding target region image patch is cropped from the original image. Simultaneously, the specific categories involved in the conflict (i.e., "truck" and "bus") are respectively constructed into formatted text descriptions (e.g., "a photo of a truck" and "a photo of a bus"). Subsequently, following the process described in steps S2432a to S2432c, the semantic similarity between the target region image patch and the "truck" text description, and the semantic similarity between it and the "bus" text description, are calculated respectively.

[0086] In step S2432e, if the category with the highest semantic similarity is a preset category among the conflicting categories, then the corresponding label is corrected to the preset category. Otherwise, the detection results and pseudo-label pairs with category conflicts are discarded.

[0087] After obtaining multiple semantic similarity scores calculated in step S2432d, a comparative analysis is performed. If, within the set of conflicting categories, there exists a predefined category (e.g., "truck") whose corresponding semantic similarity score is the highest among all conflicting categories, and this highest score has sufficient discriminative power (or satisfies a predefined confidence condition), then it is determined that external knowledge explicitly supports this category. In this case, the system uniformly corrects the labels associated with this conflicting pair to the supported predefined category. For example, if the semantic similarity of "truck" is 0.85, higher than that of "bus" (0.40), then the label is corrected to "truck". Conversely, if the semantic similarity scores between conflicting categories are very close, or both are at low levels (e.g., "truck" 0.48, "bus" 0.52), it indicates that based on the current external knowledge, neither category can be clearly distinguished or supported, and the conflicting pair is considered unreliable. In this case, the system will adopt a conservative strategy, discarding the detection results and pseudo-label pairs with category conflicts to avoid introducing uncertain noise into subsequent training loops.

[0088] By calculating and comparing the semantic similarity between visual regions and conflicting category texts, the resolution of classification ambiguities is based on quantifiable and reproducible data, avoiding the bias that may be caused by subjective or heuristic rules. Its binary decision-making mechanism of "correcting if explicitly supported and discarding if indistinguishable" can effectively clarify a large number of category confusions caused by changes in the appearance of the target domain (such as blurring in foggy weather) by using external general knowledge, significantly improving the semantic accuracy of pseudo-labels. At the same time, it can decisively filter out marginal cases that still have high uncertainty even with the help of external knowledge, preventing low-quality labels from polluting the training set.

[0089] Furthermore, step S2433, based on preset prior logic rules reflecting the co-occurrence or mutual exclusion relationships between targets, verifies the tag set after correction by the preset cross-modal semantic alignment model, including: In step S2433a, if the target category is detected as pedestrian, and there is a bicycle or motorcycle target in the spatial neighborhood, the pedestrian target category is corrected to cyclist.

[0090] In road scenarios such as autonomous driving, the category of "cyclist" is physically closely associated with vehicles such as "bicycle" or "motorcycle." If a target of category "pedestrian" is detected, and the intersection-union ratio (IU) of the bounding boxes shows that a target of category "bicycle" or "motorcycle" also exists in its spatial neighborhood (e.g., the IU of the two bounding boxes is greater than or equal to a preset threshold of 0.5), then based on the co-occurrence logic of the real world, this "pedestrian" is very likely a "cyclist" using that vehicle. In scenarios with poor visibility, such as fog, the initial model or pseudo-labels may incorrectly identify the "person + vehicle" combination as separate "person" and "vehicle" due to partial occlusion or blurred features. This rule systematically corrects such misclassified "pedestrian" targets to "cyclist" by utilizing the spatial proximity relationship between targets and category co-occurrence priors, thereby improving the logical correctness of the labels in real physical scenarios.

[0091] In step S2433b, if the target category is detected as cyclist and there are no bicycle or motorcycle category targets in its spatial neighborhood, the cyclist target category is corrected to pedestrian.

[0092] A target labeled "cyclist" is defined by its association with a mode of transportation (bicycle or motorcycle). If no targets of the "bicycle" or "motorcycle" category are detected in the spatial neighborhood of a target labeled "cyclist" (also determined by the intersection-union threshold), then the labeling is logically contradictory. In complex scenes (such as dense crowds or partial occlusion), the detection model may misclassify certain pedestrians with distinctive postures or appearances as cyclists. This rule serves as a verification method; when an isolated "cyclist" label lacking an accompanying mode of transportation is found, its category is corrected to "pedestrian," thereby eliminating logically inconsistent labels caused by model overgeneralization or misjudgment, ensuring that the labeling results conform to basic common sense in the scene.

[0093] In step S2433c, if both pedestrian and cyclist category labels exist for the same spatial location, the cyclist category label is retained and the pedestrian category label is deleted.

[0094] During object detection, due to model uncertainties or oversights in post-processing, multiple category labels may sometimes be output for the same spatial location in an image (manifested as highly overlapping bounding boxes, e.g., an intersection-union ratio greater than 0.5), which may simultaneously include "pedestrian" and "cyclist." Based on the mutual exclusivity of category definitions, an object cannot be both a "pedestrian" and a "cyclist." According to scene priors, "cyclist" describes the more specific state of "a person riding a vehicle," containing information about the "person" but with stronger scene specificity. Therefore, when such redundant conflicts occur, this rule performs a deambiguation operation, retaining the more specific and clearly defined "cyclist" category label while deleting the "pedestrian" category label. This ensures that each detected entity is ultimately assigned only one unique and most accurate category, avoiding ambiguity in the training data and improving the clarity of the supervision signal.

[0095] These rules encode common sense about co-occurrence (e.g., people and vehicles), mutual exclusion (e.g., pedestrians and cyclists), and hierarchical relationships between categories into executable judgment logic, ultimately refining the label set after semantic correction by the external model. This effectively identifies and corrects labeling inconsistencies and errors that remain after geometric matching and semantic similarity comparisons, particularly for easily confused category pairs. By introducing this knowledge-based verification layer, the logical consistency and reliability of the generated updated pseudo-label set are further improved. This ensures that the supervision signals used for the next round of training are not only semantically accurate but also conform to the physical constraints of the scene, thereby guiding the model to learn features and discrimination boundaries that are closer to real-world patterns, enhancing the rationality and robustness of the entire adaptive system's output.

[0096] The calculation process of the above detection method will be described in detail below with a specific embodiment: (1) Data collection and preprocessing Collect unlabeled target domain datasets ,in Let N represent the number of images in the dataset. The unlabeled target domain dataset... The image is scaled to 600 on the short side while maintaining the aspect ratio; common enhancements (random horizontal flipping, scale jittering, color perturbation, random cropping, etc.) can be used during the training phase to improve robustness.

[0097] (2) Learning steps target domain dataset Input into VLM (Grounding DINO) to obtain the initial set of pseudo-tags. Each pseudo-label contains the predicted bounding box coordinates. and category probability distribution vector .

[0098] for For each pseudo-label in the code, the uncertainty-aware gating mechanism performs the following operations: according to the formula Calculate its confidence level; according to the formula Calculate its information entropy; according to the formula Calculate the perceived uncertainty score. Set a score threshold T (e.g., 0.7). If S(p) ≤ T, retain the pseudo-label; otherwise, filter it out.

[0099] The model was trained using a selected set of high-quality pseudo-labels. (Preferred, YOLOv8s is used) m wheels, where m is 10 in this invention.

[0100] (3) Practical steps Use the current model For the target domain dataset Perform forward reasoning to generate a set of detection labels. (including bounding box) ,category ).

[0101] (4) Reflection Steps 1) Consistency matching Will Compared with the optimized pseudo-label set generated in the previous iteration (including bounding box) and categories If it's the first iteration, then... Matching is performed on the pseudo-label set generated during the learning process. The IoU between all bounding box pairs is calculated; if the IoU is greater than a threshold... (For example, 0.5) is considered a matching pair. Based on whether the matching pairs are of the same category, they are divided into a consistency detection set and an inconsistency detection set, where: Consistency Detection Set :satisfy ,and The tag pairs represent high-confidence tags, and the formula is: .

[0102] Inconsistency Detection Set This includes: ① (Both sets were detected, but there was a class conflict) satisfying... ,and ① There is a category conflict in the label pairs; ② Only one of the two sets has ) satisfying The label. The formula is: ; in, ; .

[0103] 2) CLIP-based correction The CLIP zero-shot classification capability is used to correct the class error of three-class sets. The core is to verify the reasonableness of the class through image embedding-text embedding similarity: CLIP Embedding Computation: For arbitrary bounding boxes via CLIP image encoder Extracting region image embedding For all categories in the target domain via CLIP text encoder Generate category text embeddings .

[0104] Cosine similarity calculation: for bounding boxes With each category Calculate similarity .

[0105] Category correction rules: right : Merge bounding boxes Retain the original category Generate a set of correction tags ; right :include and ① To Calculate the fused bounding box CLIP prediction category Only when Retain tags during generation ; ② Calculate the label bounding box CLIP prediction category Only when ( When the original tag category is retained, it is generated. . Processed calibration set and The union of two sets is given by the formula: .

[0106] Output: CLIP - Corrected tag set .

[0107] 3) Heuristic prior correction For easily confused categories in the target domain (such as "person" and "rider" in autonomous driving scenarios), corrections are made based on prior scenario knowledge. The core rules (which can be extended to other categories) are as follows: Rule 1: If a label category is "person", and its surrounding ( If a label with the category "bicycle" or "motorcycle" exists, then "person" will be corrected to "rider". Rule 2: If a tag category is "rider", and its surrounding ( If there is no "bicycle" or "motorcycle" label, then change "rider" to "person"; Rule 3: If the same goal (through (Judgment) If both the "person" and "rider" tags exist, only the "rider" tag will be retained; Output: The final optimized pseudo-tag set , used to train the detector.

[0108] 4) Use As a supervisory signal training detector n epochs, where n is 10 in this invention.

[0109] (5) Steps S2, S3, and S4 are considered as one round of iteration. Repeat steps S2, S3, and S4 for k rounds of iteration. In this invention, k is 5. Finally, the optimized target detector model is obtained.

[0110] (6) The optimized target detector model is used to detect the target domain image to be identified, and the detection and recognition results are obtained.

[0111] at last, Figure 5 The comparison of detection results between Grounding DINO, COIN, and the proposed method on the Foggy Cityscapes dataset is presented. Visually, it is clearly evident that the proposed method exhibits superior domain adaptation capabilities and outputs more accurate detection results compared to the other two methods. For example, in... Figure 5 In the first row, Grounding DINO misclassifies a train as a truck and a bus, while COIN completely misses the target; conversely, the method of this invention successfully corrects the erroneous judgment of the visual-language model, accurately identifying the target as a train. Furthermore, in Figure 5 In other complex foggy scenarios shown in the second and third rows, the method of this invention can not only more accurately locate and identify multiple targets, but also reduce the problems of false positive annotations and redundant bounding boxes in the COIN detection results.

[0112] Accordingly, a second aspect of the present invention provides an electronic device, including: at least one processor; and a memory connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to cause the at least one processor to perform the above-described efficient single-stage domain adaptive target detection method for open scenes.

[0113] Accordingly, a third aspect of the present invention provides a computer-readable storage medium having computer instructions stored thereon, which, when executed by a processor, implement the above-described efficient single-stage domain adaptive target detection method for open scenarios.

[0114] This invention aims to protect an efficient single-stage domain adaptive target detection method for open scenes, which has the following advantages: 1. By abandoning the traditional multi-stage teacher-student separation training paradigm, we innovatively propose and implement a cognitively inspired single-stage self-evolutionary closed-loop training architecture. This architecture drives the same lightweight detection model to cyclically perform cognitive iterations of "learning, practice, and reflection" on the target domain data, achieving autonomous and continuous optimization of the model. It greatly simplifies the training process, eliminates the path of error propagation and accumulation between stages, thereby improving the efficiency of domain adaptive training by an order of magnitude while maintaining high detection accuracy, and significantly reducing model complexity and computational cost, making it possible to quickly deploy and apply the model on resource-constrained edge devices. 2. To address the initial pseudo-label noise problem, an uncertainty-aware gating screening mechanism is introduced in the learning step; a breakthrough is made by fusing local confidence and global information entropy for joint evaluation, constructing a refined uncertainty-aware scoring function that can effectively identify and eliminate ambiguous or deceptive samples with "high scores but high uncertainty" that traditional methods cannot distinguish; the quality of pseudo-labels used for model supervised training is significantly improved from the source, providing a more reliable knowledge base for the subsequent iterative evolution of the model, thereby directly improving the initial performance of the model in the target domain and the stability of subsequent optimization; 3. In the reflection step, a multi-level, multi-source knowledge closed-loop correction system is constructed. Cognitive conflicts are identified through consistency matching, and powerful external visual-language models such as CLIP are introduced for zero-shot arbitration. Heuristic rules based on scenario priors are used for logical verification, forming a complete self-correction process of "discovering conflicts - adjudicating and correcting - refining logic". This effectively simulates the human ability to reflect and correct, enabling the model to continuously correct historical cognitive biases using newly acquired knowledge. This not only significantly improves the model's detection accuracy and robustness in various complex cross-domain scenarios, but also endows it with the core ability to continuously self-evolve and adapt to open and unknown environments.

[0115] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0116] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0117] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0118] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0119] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A highly efficient single-stage domain adaptive target detection method for open scenarios, characterized in that, Includes the following steps: Obtain the target domain image of the open scene to be detected; The target domain image is input into a single-stage domain adaptive model, which is a model obtained by iteratively training a cognitive closed loop that includes learning steps, practice steps and reflection steps on unlabeled target domain data. The target domain image is processed by the single-stage domain adaptive model to obtain the bounding box position and category information of each target in the target domain image.

2. The efficient single-stage domain adaptive target detection method for open scenarios according to claim 1, characterized in that, Before inputting the target domain image into the single-stage domain adaptive model, the method further includes: Retrieve unlabeled target domain data; Build or obtain the initial target detector model; Based on the target domain data, a single-stage cognitive closed-loop iterative process is executed cyclically to train and update the target detector model. The single-stage cognitive closed-loop iterative process includes learning steps, practice steps, and reflection steps in sequence. The iteration stops when the target detector model meets the convergence condition, thus obtaining the single-stage domain adaptive model.

3. The efficient single-stage domain adaptive target detection method for open scenarios according to claim 2, characterized in that, The iterative process of executing a single-stage cognitive closed-loop process to train and update the target detector model includes: In the first iteration, a pseudo-label set is generated based on the target domain data and the prediction-trained visual-language large model with open vocabulary detection capability. The pseudo-label set contains the predicted target bounding box coordinates and their corresponding category probability distributions. In the current iteration, the currently held pseudo-label set is used as a supervision signal to supervise the training of the target detector model, thus completing the learning step. Using the target detector model updated through the learning steps, forward reasoning is performed on the unlabeled target domain data to generate the detection result set for the current round, thus completing the practical steps. The detection result set of the current round is compared and verified with the pseudo-label set used as input to the learning step of the current round, and the detection results are corrected according to the verification results to generate an updated pseudo-label set, thus completing the reflection step. The updated pseudo-label set is used as the input for the next iteration, replacing the currently held pseudo-label set, and the learning step, the practice step, and the reflection step are repeated.

4. The efficient single-stage domain adaptive target detection method for open scenarios according to claim 3, characterized in that, Before using the currently held pseudo-label set as a supervision signal to perform supervised training on the target detector model, the method further includes: For each pseudo-label in the currently held pseudo-label set, based on the category probability distribution it contains, calculate the local confidence level reflecting the category certainty and the global information entropy reflecting the degree of dispersion of the probability distribution; By fusing the local confidence score with the global information entropy, the uncertainty perception score of the pseudo-label is calculated; The uncertainty perception score is compared with a preset quality threshold, and pseudo-labels with scores better than the quality threshold are retained to form a high-quality pseudo-label set for supervised training.

5. The efficient single-stage domain adaptive target detection method for open scenarios according to claim 4, characterized in that, The uncertainty perception score of the pseudo-label The calculation formula is: ； ； ； in, The global information entropy, The local confidence level is... Let be the probability value of the i-th category of the pseudo-label, and s be the preset number of categories in the target domain. For uncertain adjustment term weights, This is the sensitivity scaling factor. The slope parameter of the Sigmoid function. The weight of the confidence penalty term. This is the confidence decay index.

6. The efficient single-stage domain adaptive target detection method for open scenes according to any one of claims 3-5, characterized in that, The step of comparing and verifying the detection result set of the current round with the pseudo-label set used as input to the learning step of the current round, and correcting the detection results based on the verification results, includes: The detection result set of the current round is matched with the pseudo-label set for spatial and category consistency. Based on the overlap of the bounding boxes and the consistency of the category labels, the matching results are divided into a consistent detection subset and an inconsistent detection subset. For the inconsistency detection subset, a preset cross-modal semantic alignment model is invoked to calculate the semantic similarity between the target region image and the candidate category text description in the inconsistency detection subset, and the conflicting category labels are corrected based on the semantic similarity. Based on the preset prior logic rules that reflect the co-occurrence or mutual exclusion relationship between target scenes, the tag set after being corrected by the preset cross-modal semantic alignment model is verified to obtain the updated pseudo-tag set.

7. The efficient single-stage domain adaptive target detection method for open scenarios according to claim 6, characterized in that, The step of performing spatial and category consistency matching between the current round's detection result set and the pseudo-label set includes: Calculate the intersection-union ratio (IUU) between the bounding box of each detection result in the current round of the detection result set and the bounding box of each pseudo-label in the pseudo-label set; The detection results with the same category label and the intersection-union ratio greater than the preset overlap threshold are divided into the consistent detection subset along with the pseudo-label pairs; Detection results or pseudo-labels that have an intersection-union ratio greater than the preset overlap threshold but have different category labels, or that fail to reach the preset overlap threshold with any element in the other set, are classified into the inconsistency detection subset.

8. The efficient single-stage domain adaptive target detection method for open scenarios according to claim 6, characterized in that, The calculation of the semantic similarity between the target region image and the candidate category text description in the inconsistency detection subset includes: The image encoder of the preset cross-modal semantic alignment model extracts the image embedding vector of the target region image; The text encoder of the preset cross-modal semantic alignment model extracts the text embedding vector of the candidate category text description; The cosine similarity between the image embedding vector and the text embedding vector is calculated as the semantic similarity.

9. The efficient single-stage domain adaptive target detection method for open scenarios according to claim 6, characterized in that, The correction of conflicting category labels based on the semantic similarity includes: For the detection results and pseudo-label pairs with category conflicts in the inconsistency detection subset, calculate the semantic similarity between the corresponding target region image and the text description corresponding to the conflict category; If the category with the highest semantic similarity is a preset category among the conflicting categories, then the corresponding label is corrected to the preset category; otherwise, the detection results and pseudo-label pairs with category conflicts are discarded.

10. The efficient single-stage domain adaptive target detection method for open scenarios according to claim 6, characterized in that, The preset prior logic rules, based on the co-occurrence or mutual exclusion relationships between targets, are used to validate the tag set after correction by the preset cross-modal semantic alignment model, including: If a target category of pedestrian is detected, and there are bicycle or motorcycle targets in the vicinity of the pedestrian, the pedestrian target category is changed to cyclist. If a target category of cyclist is detected, and there are no bicycle or motorcycle category targets in its spatial neighborhood, the cyclist target category is corrected to pedestrian. If both pedestrian and cyclist category labels exist for the same spatial location, the cyclist category label will be retained and the pedestrian category label will be deleted.

Images