A semi-supervised remote sensing small target detection method based on target-level discriminative prior transfer

Abstract

This invention relates to the technical field of target detection and discloses a semi-supervised remote sensing small target detection method based on target-level discriminative prior transfer. First, a teacher model is used to generate candidate target-level discriminative priors on unlabeled images. Second, based on the candidate target-level discriminative priors, a joint reliability judgment is performed from two dimensions: category determinism and location stability. Sample-level training weights are assigned based on the target score to generate a reliable prior set. Finally, a student model performs supervised training and unsupervised prior transfer training according to the input sample type and the reliable prior set, respectively. The trained student model is then used to perform target detection on the remote sensing image to be detected.

Description

Technical Field

[0001] This invention belongs to the technical field of target detection, specifically relating to a semi-supervised remote sensing small target detection method based on target-level discriminative prior transfer. Background Technology

[0002] In recent years, deep learning technology has significantly promoted the development of target detection methods. Models such as Faster R-CNN, Detection Transformer (DETR), Denoising Anchor Box Distillation Detector (DINO), and Real-Time Detection Transformer (RT-DETR) have achieved excellent performance in the field of natural images and are gradually being applied to remote sensing target detection tasks. However, compared with natural images, remote sensing images usually have characteristics such as top-down perspective, complex background, significant scale changes, a high proportion of small targets, and weak inter-class differences. This makes existing detection methods face greater challenges in recognition and localization in remote sensing scenarios.

[0003] In practical applications, the cost of labeling remote sensing images is high, especially in high-resolution remote sensing images. Small targets often have characteristics such as blurred boundaries, small scale, and sparse distribution, and their accurate labeling usually requires a lot of manual time and professional experience. In contrast, unlabeled remote sensing images are easier to obtain. Therefore, under the condition of a small amount of labeled data and a large amount of unlabeled data, how to effectively utilize the potential supervision information in unlabeled samples has become an important research problem in the field of remote sensing target detection.

[0004] Semi-supervised object detection offers an effective way to alleviate label dependence. Current mainstream methods typically employ a teacher-student framework, where a teacher model generates pseudo-labels on unlabeled images, and a student model is jointly trained using real labels and pseudo-labels. For example, the Unbiased Teacher method alleviates pseudo-label bias through unbiased learning, the Soft Teacher model improves the utilization efficiency of unlabeled samples through soft labels and weak-strong enhancement strategies, and the Dense Teacher model further enhances pseudo-label learning capabilities from a dense supervision perspective. Although these methods have achieved good results in the natural image domain, their performance still highly depends on the quality of the teacher's output. When there are many false detections, classification errors, or localization biases in the pseudo-labels, the student model is easily affected by noisy supervision, leading to performance limitations. This problem is even more prominent in remote sensing small target detection scenarios: due to the small target size, weak texture, and strong background interference, the classification confidence relying solely on the teacher model's output often fails to accurately reflect the true usability of candidate targets.

[0005] In recent years, the Grounding DINO (Location-Guided Denoising Anchor Box Improved Detection Transformer) has demonstrated strong cross-modal generalization capabilities in open-set target detection and pointer expression detection tasks by combining the DINO detector with localization-guided grounded pre-training. Compared to traditional closed-set teacher detectors, Grounding DINO not only has stronger target localization capabilities but also enhances visual semantic representations with textual cues, providing new possibilities for generating more discriminative candidate targets on the teacher side. However, in low-label remote sensing small target detection scenarios, even with a stronger teacher model, directly using teacher output as pseudo-labels may still be affected by class ambiguity, localization jitter, and false detections in complex backgrounds. Therefore, identifying and retaining truly reliable candidate targets from teacher output and transferring them to the student model in a reasonable manner is key to improving the performance of semi-supervised remote sensing small target detection. Summary of the Invention

[0006] This invention provides a semi-supervised remote sensing small target detection method based on target-level discriminative prior transfer, in order to solve the technical problems of large pseudo-supervisory noise, unstable positioning, and insufficient utilization efficiency of unlabeled samples in remote sensing small target detection under low labeling conditions.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A semi-supervised remote sensing small target detection method based on target-level discriminative prior transfer is proposed. First, a teacher model is used to generate candidate target-level discriminative priors on unlabeled images. Second, based on the candidate target-level discriminative priors, a joint reliability judgment is performed from two dimensions: category certainty and location stability. Sample-level training weights are assigned based on the target score to generate a reliable prior set. Finally, a student model performs supervised training and unsupervised prior transfer training according to the input sample type and the reliable prior set, respectively. The trained student model is then used to perform target detection on the remote sensing image to be detected. Furthermore, for any unlabeled image The candidate target-level discriminative prior set output by the teacher model Represented as in, Indicates the number of candidate targets. Indicates the first Predicted bounding boxes of candidate targets, This represents the corresponding category probability distribution. This represents the corresponding goal score, indicating the confidence level that the candidate goal belongs to the prospective goal. Indicates the total number of categories; When performing joint reliability determination, firstly, a category determinism score is calculated based on the category probability distribution, and then a category availability mask is generated. Simultaneously, a location stability score is calculated based on the predicted bounding box, and then a location availability mask is generated. Finally, a joint availability mask is generated based on the category availability mask and the location availability mask. Then, based on the target score, sample-level training weights are generated. Then the image Corresponding reliable prior set Represented as The reliable prior set corresponding to the entire unlabeled sample set Represented as in, Indicates the number of unlabeled samples. This indicates the number of candidate targets.

[0008] Furthermore, based on the category determinism threshold Define category availability mask for in, Represents the information entropy function. Indicates the category certainty score; Based on positioning stability threshold Define the location availability mask for in, This represents the positioning stability score. Indicates the intersection-union ratio of the bounding boxes. Indicates unlabeled images Apply geometric transformation The augmented image is then obtained, and the teacher model's predicted bounding box is located on the augmented image. Representing geometric transformations inverse transform; Calculate the joint availability mask using the following formula. , in, Indicates the first Whether the candidate target is retained.

[0009] Furthermore, the sample-level training weights are calculated using the following formula. , in, Indicates the first The sample-level training weights corresponding to each candidate target.

[0010] Furthermore, during the training of the student model, supervised learning is performed directly based on the real annotations for labeled samples; For unlabeled samples, the parameters corresponding to the candidate targets are first generated using the student model, and then unsupervised prior transfer training is performed in combination with a reliable prior set.

[0011] Furthermore, for labeled samples The student model is directly based on real annotations. Supervised learning is performed, and its loss function is... Defined as in, This indicates supervised classification loss. This indicates the supervised bounding box regression loss. Indicates supervised regression loss The weighting coefficients.

[0012] Furthermore, for unlabeled samples The student model output is represented as in, , and These represent the student model in the first... Predicted bounding boxes, class probability distributions, and target scores for each candidate target. Indicates the number of candidate targets; First, for each sample from the reliable prior set, a one-to-one matching strategy based on maximizing the bounding box IoU is adopted to find the sample in the unlabeled sample set that is closest to the geometric position of the candidate target, thereby achieving sample matching. These samples are then used together for training the student model. Each candidate target is matched with at most one sample from the reliable prior set. Secondly, the unsupervised migration loss is calculated using the following formula. ; in, Indicates the Kullback-Leibler divergence. and These represent the predicted bounding boxes output by the student model and the teacher model, respectively. This represents the target score output by the teacher model. This represents the target score output by the student model. Indicates the weight of location migration loss. Indicates the weight of the target migration loss. , , These represent the category migration loss, location migration loss, and destination migration loss, respectively. Represents the joint availability mask. This indicates the generation of sample-level training weights.

[0013] Furthermore, the teacher model is configured as a localization-guided denoising anchor box improved detection transformer (GroundingDINO), and the student model is configured as a real-time detection transformer (RT-DETR).

[0014] Compared with the prior art, the beneficial effects of the present invention are: 1. A target-level discriminative prior reliability transfer framework for remote sensing small target detection is proposed, which establishes a target-level knowledge transfer mechanism between the Grounding DINO teacher model of the localization-guided denoising anchor frame improved detection transformer and the RT-DETR student model of the real-time detection transformer.

[0015] 2. Design a joint determination mechanism for category determinism and positioning stability to screen high-quality candidate priors from both semantic reliability and geometric stability aspects, thereby reducing the negative impact of noise supervision propagation.

[0016] 3. A target-aware sample-level weight allocation strategy is proposed to achieve hierarchical modeling from "whether the candidate is available" to "the strength of the candidate's effect", thereby improving the stability and effectiveness of the training process for unlabeled samples.

[0017] 4. Comparative experiments on the DIOR dataset show that the proposed method outperforms many mainstream fully supervised and semi-supervised methods, achieving an mAP of 76.9%, thus verifying the effectiveness of the proposed method in remote sensing target detection.

[0018] 5. Ablation experiments further demonstrate that category determinism, localization stability constraints, and target perception weighting strategies all make significant contributions to improving model performance.

[0019] Overall, the method of this invention is designed around unsupervised signal quality control, and improves the accuracy, stability and interpretability of semi-supervised remote sensing small target detection without changing the main structure of the student detector. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall framework of the present invention. Detailed Implementation

[0021] To make the technical means, creative features, objectives and effects of this invention easier to understand, the following embodiments, in conjunction with the accompanying drawings, specifically illustrate the semi-supervised remote sensing small target detection method based on target-level discrimination prior transfer of this invention. It should be noted that the description of these embodiments is for the purpose of helping to understand this invention, but does not constitute a limitation of this invention.

[0022] This invention proposes a semi-supervised remote sensing small target detection method based on target-level discriminative prior transfer. It employs a localization-guided denoising anchor box improved detection transformer (Grounding DINO) as the fixed teacher model and a real-time detection transformer (RT-DETR) as the student model. Instead of directly treating the teacher's predictions as pseudo-labels, it uniformly represents them as candidate target-level discriminative priors. Based on this, a joint reliability assessment of the candidate priors is performed from two dimensions: category determinism and localization stability, retaining only semantically clear and geometrically stable candidate targets. Furthermore, the target-specific scores output by the teacher are used to assign sample-level training weights to the reliable priors, guiding the student model to differentially absorb high-quality target-level knowledge. Through this design, the utilization of unlabeled samples is transformed from "direct pseudo-label supervision" to "reliable prior screening and weighted transfer," effectively reducing the negative transfer risk caused by noise supervision propagation.

[0023] To fully utilize the large number of unlabeled remote sensing images under low-label conditions, this invention adopts a fixed teacher-student framework and designs the model around three stages: candidate target-level discriminative prior generation on the teacher side, prior reliability screening and sample-level weighting, and supervised-unsupervised joint optimization on the student side, thus completing the construction of the recognition model R2T-Det. The overall process of the proposed method is as follows: Figure 1 As shown, the process mainly includes three stages: First, a fixed localization-guided denoising anchor box improves the detection transformer GroundingDINO teacher model to generate candidate target-level discriminative priors on unlabeled images; second, the candidate priors are jointly evaluated for reliability based on both category determinism and localization stability, and sample-level training weights are assigned based on the target score; finally, the student model performs supervised training and unsupervised prior transfer training according to the input sample type, thereby achieving robust utilization of unlabeled data. Figure 1In this module, module A corresponds to the generation of teacher candidate priors, module B corresponds to the reliability determination and weighting of candidate priors, and module C corresponds to the joint optimization process of the student model.

[0024] Specifically as follows: I. Problem Definition and Overall Framework To provide a consistent description of the method of this invention, let the labeled training set be... in, Indicates the first Zhang's labeled remote sensing image, Representation and Image Corresponding target annotation, This indicates the number of labeled samples.

[0025] For images The annotation is denoted as in, Representing an image Number of target annotations in the data. Show the first The true bounding box of each target This indicates the corresponding category label.

[0026] Let the unlabeled training set be in, Indicates the first Zhang's unlabeled remote sensing image, This indicates that the number of samples was not labeled.

[0027] Teacher model denoted as The student model is denoted as In this invention, To improve the detection transformer Grounding DINO for positioning and guiding noise reduction anchor frames, This is RT-DETR. The teacher model is first pre-trained on labeled data, then its parameters are fixed and it is used only for candidate prior generation on unlabeled images; the student model selects different optimization branches based on the type of input samples during the training phase. The overall training objective can be represented as... in, This represents the supervised loss on the labeled samples. Indicates based on a reliable prior set The unsupervised transfer loss is used. For labeled samples, the student model learns directly using the real labels; for unlabeled samples, the student model uses reliable priors generated by the teacher model and filtered for transfer learning.

[0028] II. Teacher Model Construction and Candidate Prior Generation This section corresponds to Figure 1 Module A of the demonstrated R2T-Det recognition model aims to generate candidate target-level discriminative priors on unlabeled images for subsequent reliability assessment and transfer learning. Unlike traditional semi-supervised detection methods that directly use teacher predictions as pseudo-labels, this invention first uniformly represents the teacher output as candidate target-level discriminative priors to facilitate consistent reliability screening and weighted modeling in the future.

[0029] For any unlabeled image The set of candidate targets output by the teacher model is represented as in, Indicating the teacher model in the image The candidate set generated above, Indicates the number of candidate targets. Indicates the first Predicted bounding boxes of candidate targets, This represents the corresponding category probability distribution. The corresponding target score represents the confidence that the candidate target belongs to the foreground target rather than the background. In this invention, it can be obtained from the foreground probability output by the detection head of the teacher model, or calculated from the aggregate probability of non-background categories in the category probability distribution. This indicates the total number of categories.

[0030] Further, define the image The Middle The prior for target-level discrimination of candidate targets is: in, It also includes location information. Category distribution information With target information Therefore, the teacher model output is no longer simply regarded as a single pseudo-label, but is uniformly modeled as a candidate discriminant prior that includes semantic and geometric attributes, laying the foundation for subsequent reliability assessment and differentiated transfer.

[0031] III. Determination and Weighting of Candidate Prior Reliability This section corresponds to Figure 1The demonstration shows module B of the R2T-Det recognition model. Considering that candidate results generated by the teacher model in complex remote sensing contexts may still suffer from semantic ambiguity and localization jitter, this invention performs a joint reliability assessment of the candidate target-level discriminative prior from two dimensions: category certainty and localization stability, and further utilizes target-specific scores for sample-level weighting. Through this process, a reliable prior set that can be used for unsupervised transfer learning is finally obtained.

[0032] For images To obtain candidate targets, this invention needs to obtain two results: first, a joint mask indicating whether the candidate targets are usable. Secondly, the corresponding training weights. Based on this, the image The reliable prior set can be represented as The reliable prior set over the entire unlabeled data is represented as in, This represents the set of unlabeled samples that can be used for unsupervised transfer training.

[0033] 3.1 Classification Determination Category certainty measures whether candidate targets have a clear semantic category affiliation. This applies to the category distribution of candidate targets. Its category certainty score is defined as in, Represents the information entropy function. Indicates the total number of categories. This represents the category certainty score. The higher the score, the more concentrated the category distribution and the lower the semantic uncertainty of the candidate target.

[0034] Thus, due to the weak texture, small scale, and strong background interference of small remote sensing targets, some false positive candidates may have a high probability in one category, but the overall distribution is still not sharp enough. The conventional maximum class probability only reflects the local confidence level of the optimal class and is difficult to characterize the concentration of the entire category distribution. Therefore, this invention uses normalized entropy to calculate category certainty, which can measure semantic uncertainty from the perspective of the overall distribution. It is more suitable for screening out candidate targets with ambiguous category boundaries and can better identify such candidates that "look confident but are actually ambiguous".

[0035] Based on category deterministic threshold Define the category availability mask as in, This indicates the availability determination result of the candidate target at the category level.

[0036] 3.2 Positioning Stability Determination Localization stability measures whether candidate targets exhibit a consistent spatial response at the geometric level. For unlabeled images Apply geometric transformation The resulting enhanced image The corresponding predicted bounding box of the teacher model on the augmented image is denoted as to transform it by inverse transformation After mapping back to the original coordinate system, the positioning stability score is defined as follows: in, This represents the positioning stability score. This represents the intersection-union ratio (IoU) of the bounding boxes. The higher the score, the more consistent the geometric position of the candidate target is between the original image and the enhanced view, and the more stable the localization.

[0037] Based on positioning stability threshold Define the location availability mask as in, This indicates the availability determination result of the candidate target at the positioning level.

[0038] 3.3 Joint Reliability Screening Only after satisfying the criteria at both the category and location levels can a candidate target qualify for migration. Its joint availability mask is defined as follows: in, Indicates the first Whether a candidate target is retained. Only when a candidate target simultaneously meets the conditions of clear category and stable location is it considered a reliable prior for subsequent unsupervised transfer training.

[0039] 3.4 Goal-oriented perception weighting Even after joint reliability screening, the supervisory value among candidate targets may still differ. To further characterize the relative contributions of different reliable priors to student model training, this invention constructs sample-level training weights based on the target scores output by the teacher model, defined as follows: in, Indicates the first The training weights corresponding to each candidate target The target score is output by the teacher model. Thus, the candidate prior not only has a discrete decision result on whether to participate in training, but also has a continuously adjustable representation of the supervision intensity, thereby realizing hierarchical modeling of supervision information for unlabeled samples, that is, realizing two-stage reliable prior modeling from "candidate availability determination" to "supervision contribution allocation".

[0040] IV. Joint Optimization of Student Models This section corresponds to Figure 1 The demonstrated recognition model, R2T-Det, uses module C. The student model simultaneously utilizes real-world supervision information from labeled samples and reliable priors from unlabeled samples for joint optimization. The training process consists of two parts: a supervised branch and an unsupervised prior transfer branch. For any input sample, the optimization objective can be written as... in, , and These represent the weight coefficients of supervised regression loss, location migration loss, and target migration loss, respectively.

[0041] 4.1 Supervised training For labeled samples The student model is directly based on real annotations. Supervised learning is performed, and its loss function is defined as follows: in, This indicates supervised classification loss. This represents the supervised bounding box regression loss.

[0042] 4.2 Unsupervised Prior Transfer For unlabeled samples The student model output is represented as in, , and These represent the student model in the first... Boundary box prediction, class probability distribution, and targetability output for each candidate target.

[0043] First, for each sample from the reliable prior set, a one-to-one matching strategy based on maximizing the bounding box IoU is adopted to find the sample in the unlabeled sample set that is closest to the geometric position of the candidate target, thereby achieving sample matching. These samples are then used together for training the student model. Each candidate target is matched with at most one sample from the reliable prior set. Then, the unsupervised migration loss is calculated using the following formula; Based on a reliable prior set, the unsupervised transfer loss on unlabeled samples is defined as: in, Indicates the weight of location migration loss. This represents the weight of the target migration loss.

[0044] Unsupervised prior transfer aims to enable student models to simultaneously inherit semantic discriminative information (category distribution), spatial location information (bounding boxes), and foreground saliency information (targetability) from the reliable priors of teacher models. Therefore, this invention decomposes unsupervised loss into three parts: category transfer loss, location transfer loss, and targetability transfer loss.

[0045] I. Category Migration Loss The Kullback-Leibler divergence constraint is used to align the student model's class distribution with the teacher model's reliable prior distribution. in, This represents the Kullback-Leibler divergence.

[0046] II. Location migration loss By employing bounding box constraints based on the intersection-union ratio (IoU), the predicted bounding boxes of the student model are made to approximate the reliable bounding boxes of the teacher model. in, and These represent the bounding box predictions for the student model and the teacher model, respectively.

[0047] III. Loss of Targeted Migration The goal-oriented scores output by the teacher model are used to supervise the goal-oriented predictions of the student model, defined as follows: in, This represents the target score of the teacher model. This represents the target output of the student model.

[0048] Through the above design, the student model establishes basic detection capabilities using a small number of labeled samples, and fully absorbs potential supervision information from unlabeled data by leveraging target-level discrimination priors that have undergone reliability screening and sample-level weighting. This improves the accuracy, positioning quality, and training stability of remote sensing small target detection under low-label conditions.

[0049] To comprehensively evaluate the applicability of the proposed method in remote sensing scenarios of varying complexity, this invention selects the publicly available remote sensing target detection dataset DIOR for experiments.

[0050] 1. Dataset DIOR is a representative large-scale public dataset in the field of remote sensing target detection. It includes 20 target categories such as aircraft, ships, vehicles, bridges, airports, and oil tanks. It features rich scenes, complex backgrounds, and a wide range of target scale variations, and can well reflect the multi-class, multi-scale, and complex background features in real remote sensing images. Due to the diverse category distribution and scene types in DIOR, it is more suitable for evaluating the overall detection performance and generalization ability of models in complex remote sensing environments.

[0051] To meet the requirements of the semi-supervised detection task, the experiment adopted a training mode of "a small number of labeled samples + unlabeled samples". The labeled set was used for pre-training of the teacher model and supervised branch training of the student model, while the unlabeled set was used for candidate prior generation on the teacher side and unsupervised transfer learning on the student side.

[0052] 2 Evaluation Indicators Based on the comparative experiments conducted on the DIOR dataset according to this invention, category accuracy (AP) and mean overall accuracy (mAP) were used as evaluation metrics. Specifically, category AP measures the model's detection performance across different target categories such as aircraft, airports, bridges, ships, and vehicles; mAP comprehensively reflects the model's overall detection accuracy across all categories. Unless otherwise specified, all experimental results were statistically analyzed under the same dataset partitioning and evaluation protocol to ensure fairness and comparability in comparisons between different methods.

[0053] 3. Model Implementation and Training Details The experiments of this invention were conducted on an AMD 1950X 3.4 GHz processor, 32 GB of memory, and an NVIDIA GeForce RTX 4080 graphics card. The software environment consisted of Ubuntu 24.04, Python 3.10, CUDA 12.2, PyTorch 2.0.1, and the corresponding version of Torchvision.

[0054] The teacher model uses Grounding DINO-T, the visual backbone network is Swin-T, and the text encoder uses BERT-base. This teacher model is first trained on labeled data, then its parameters are fixed and used only for generating prior information for candidate object-level discrimination on unlabeled images. The RT-DETR student model uses ResNet-50 as its backbone network. To further analyze the scalability of the method and compare it with different fully supervised and semi-supervised methods, the input images are uniformly scaled to 1024×1024.

[0055] The text prompts for the teacher model's inference phase are constructed by concatenating all category names in the dataset in a fixed order to ensure consistency of semantic prompts on the teacher's end across different experimental settings. The student model is trained using the AdamW optimizer, with the learning rate, weight decay, number of training epochs, and batch size set according to the training configuration of the corresponding baseline model. For data augmentation, labeled samples undergo conventional augmentations such as random flipping, scaling, and color perturbation; unlabeled samples receive additional lightweight geometric augmentations beyond the base view to calculate the consistency of teacher predictions across different views. A category determinism threshold is also applied. Set to 0.7, positioning stability threshold Set to 0.6, the weight for location migration loss. Set to 1.0. Changing the parameters will not alter the principle of this invention. It should be noted that the method of this invention maintains consistency with the corresponding RT-DETR baseline in terms of student model structure, parameter size, and computational complexity. Therefore, the performance improvement mainly stems from the improved utilization of unlabeled samples, rather than the introduction of additional model complexity. This setting helps to more fairly verify the effectiveness of the reliable prior transfer mechanism itself.

[0056] 4. Comparison of experimental results To verify the target detection performance of the method of the present invention in complex remote sensing scenarios, comparative experiments were conducted on the DIOR dataset using supervised detection methods, semi-supervised detection methods, and representative models of different scales. The AP of each category and the overall mAP were used as evaluation indicators. The specific results are shown in Table 1.

[0057] Table 1. Comparative experimental results on the DIOR dataset. As shown in Table 1, the R2T-Det recognition model proposed in this invention achieved 76.9% mAP on the DIOR dataset, which is better than many existing supervised detection methods and semi-supervised detection baselines, demonstrating strong comprehensive detection capabilities. Compared to the semi-supervised baseline, R2T-Det achieves performance improvements of 25.0, 11.0, 9.6, and 6.7 percentage points over the purely supervised method Supervised Only, the soft teacher model SoftTeacher, the semi-supervised detection transformer Semi DETR, and the enhanced mean teacher method EMT, respectively. Compared to some representative supervised detectors, R2T-Det also outperforms Faster R-CNN, YOLOv3 (third version of the single-stage object detection network), AFADet (adaptive feature aggregation detector), RT-DETR-1 (large real-time detection transformer), and RMRN-DETR, with improvements of 22.8, 19.8, 10.8, 11.4, and 8.5 percentage points, respectively. It is only slightly higher than the 76.2% performance of the visual-language prior remote sensing detector VLPRSDet, still achieving a 0.7 percentage point advantage. These results demonstrate that the method of this invention possesses good detection accuracy and competitiveness in the DIOR remote sensing object detection task.

[0058] Further analysis of the category detection results shows that the proposed method achieves optimal or near-optimal performance across multiple categories. For example, R2T-Det demonstrates significant advantages in categories such as airplanes, airports, basketball courts, highway service areas, athletic fields, ships, train stations, and vehicles, achieving APs of 89.4, 79.8, 92.6, 89.9, 84.6, 92.1, 91.7, and 72.6, respectively. Simultaneously, the method also achieves high accuracy in categories such as baseball fields, bridges, chimneys, and golf courses, indicating that it has good adaptability to categories with large scale variations, complex backgrounds, and sparse target distribution in remote sensing scenes.

[0059] The results show that R2T-Det provides more significant improvements for categories with obvious structural features or rich contextual information, such as basketball courts, highway service areas, ships, and train stations. This indicates that the method of this invention, by improving the utilization of unlabeled samples, can provide more reliable supervision signals for student models, thereby enhancing their ability to discriminate and locate target areas. For categories such as ports and overpasses, although the improvement is relatively limited, the overall performance remains at a high level, indicating that the method has good robustness in complex remote sensing contexts.

[0060] The results above show that the performance improvement of the proposed method on the DIOR dataset mainly stems from the effective mining of information from unlabeled samples, rather than relying on additional model structure expansion. Since the student model backbone in the experiments still uses ResNet-50, the improved detection performance more directly reflects the effectiveness of the proposed method in semi-supervised remote sensing target detection scenarios.

[0061] The following conclusions can be drawn from the results of the comparative experiment.

[0062] First, the proposed method significantly outperforms several existing semi-supervised detection baselines on the DIOR dataset, demonstrating its strong effectiveness in remote sensing target detection tasks. In particular, compared to the Soft Teacher model, the Semi DETR semi-supervised detection transformer, and the Enhanced Mean Teacher (EMT) method, R2T-Det achieves a significant improvement in overall mAP, indicating that the proposed method can more fully utilize the effective information in unlabeled samples.

[0063] Second, the proposed method achieved high detection accuracy across multiple categories, demonstrating its strong category generalization ability. Particularly in categories such as airplanes, basketball courts, highway service areas, ships, and train stations, R2T-Det achieved outstanding AP performance, reflecting its ability to maintain strong detection capabilities even in complex scenes, dense backgrounds, and under varying scale conditions.

[0064] Third, the proposed method not only outperforms semi-supervised baselines but also has competitive advantages over several representative supervised detection methods, demonstrating that the method of this invention can still achieve high-quality detection under low-labeling conditions. This result shows that by improving the unlabeled sample utilization mechanism, model performance can be effectively improved without relying on additional manual annotation, thereby enhancing the practical application value of semi-supervised remote sensing target detection methods.

[0065] Fourth, the performance improvement of the method in this invention mainly stems from the enhancement of reliable supervision information within the method itself, rather than the introduction of additional model complexity. Since no additional extensions were made to the student model detection structure in the experiments, the performance gains shown in Table 1 demonstrate that the method in this invention achieves superior detection results while maintaining a basically consistent model framework.

[0066] 4.5 Ablation Experiment 4.5.1 Ablation Analysis of the Module First, to verify the effectiveness of the three-part design of category determinism, localization stability, and target perception weighting, this invention gradually introduces each module based on the DETR student model trained using only labeled data. The experimental results are shown in Table 2.

[0067] Table 2 Module Ablation Experiment Results As shown in Table 2, both category determinism and localization stability can individually improve performance, indicating that selecting teacher candidates from both semantic reliability and geometric stability perspectives is necessary. In comparison, localization stability has a more significant impact on mAP75, demonstrating its stronger effect on improving bounding box quality under strict localization conditions. Furthermore, when both decision mechanisms are introduced simultaneously, the model performance improves even more significantly, indicating their strong complementarity in remote sensing small target scenarios. Finally, by adding target-aware weighting to the joint selection, the model achieves even better results, demonstrating that even within reliable candidates, differences in supervisory value exist between different priors, making differentiated weighting necessary.

[0068] 4.5.2 Ablation experiments at different labeling ratios To further verify the effectiveness of the proposed method under different labeled resource conditions, this invention sets different labeled ratios on the DIOR dataset and conducts comparative experiments on the baseline model and the proposed method. By controlling the proportion of labeled samples and treating the remaining training samples as unlabeled data, the ability of the proposed method to utilize unlabeled samples under different supervision intensities can be evaluated. The experimental results are shown in Table 3.

[0069] Table 3 Ablation experiment results under different labeling ratios As shown in Table 3, the detection performance of both the baseline model and the proposed method gradually improves with increasing annotation ratio, indicating that more comprehensive manual annotation enhances the direct supervision capability of the student model. Meanwhile, the proposed method outperforms the baseline model at all annotation ratios, and typically exhibits more significant performance gains at lower annotation ratios. This demonstrates that the reliable prior transfer mechanism proposed in this invention can more effectively mine potential supervisory information in unlabeled samples when annotation resources are limited, thereby mitigating the performance degradation caused by insufficient annotation.

[0070] Furthermore, the proposed method typically achieves a higher improvement in mAP75 than mAP, indicating that it not only improves overall detection accuracy but also consistently enhances bounding box localization quality. This aligns with the localization stability determination mechanism introduced in this invention, further validating the effectiveness of this design for remote sensing small target detection tasks.

[0071] 4.5.3 Ablation Analysis of Parameters Table 4. Category Determinism Thresholds Parameter ablation experiment As can be seen from Table 4, with Increasing the threshold from 0.5 to 0.7 gradually improved model performance, indicating that appropriately raising the category certainty threshold helps filter semantically ambiguous candidate targets. When the threshold was further increased to 0.8 and 0.9, performance began to decline, indicating that excessively high thresholds could mistakenly remove some usable candidates, reducing the utilization of unlabeled data. Overall, The effect is best when the value is 0.7.

[0072] Table 5 Positioning stability thresholds Parameter ablation experiment As can be seen from Table 5, appropriately increasing... This helps suppress geometrically unstable candidates, especially significantly improving mAP75. When the threshold is too low, some candidates with significant positional drift are preserved; when the threshold is too high, some valuable prior knowledge is lost. Experimental results show that... A good balance was achieved between accuracy and sample retention rate when the value was 0.6.

[0073] Table 6 Targeted Loss Weights Parameter ablation experiment As shown in Table 6, appropriately introducing target-based transfer helps enhance the student model's ability to discriminate foreground regions, but its weight is not necessarily better the larger it is. Smaller weights are preferable. It is difficult to fully leverage the role of targeted supervision, while excessive supervision... This weakens the effects of category and position transfer, leading to an overall performance decrease. In the experiment, Performance is best when the value is 1.0.

[0074] 4.5.4 Comparison Experiment of Detection Performance and Computational Complexity To further verify the comprehensive performance of the method of this invention in terms of detection performance and computational complexity, RT-DETR was selected as a comparison method. The model performance was analyzed from four indicators: mAP50, mAP50-95, Precision, and GFLOPs. The experimental results are shown in Table 7.

[0075] Table 7 Comparison of Detection Performance and Computational Complexity As shown in Table 7, the method of this invention achieves detection performance comparable to or even better than the comparative methods while maintaining low computational complexity. Specifically, RT-DETR achieves 85.38% and 62.90% in mAP50 and mAP50-95, respectively, while the method of this invention achieves 85.9% and 63.0%, respectively, representing improvements of 0.52 and 0.10 percentage points over RT-DETR. This indicates that the method of this invention has certain advantages in overall detection accuracy and comprehensive detection capability under different cross-union ratio thresholds.

[0076] Regarding the precision metric, the RT-DETR result was 86.21%, while the method of this invention reached 87.8%, an improvement of 1.59 percentage points. This result demonstrates that the method of this invention outperforms the method in terms of target prediction accuracy, effectively reducing false detections and improving the reliability of detection results during the detection process.

[0077] In terms of computational complexity, RT-DETR has a GFLOPs of 103.5, while the method of this invention has a GFLOPs of 101.2, a reduction of 2.3 GFLOPs. This indicates that the method of this invention improves detection performance without introducing additional high computational overhead, and instead reduces model complexity to a certain extent, thus achieving better computational efficiency.

[0078] In summary, the results show that the method of this invention outperforms the real-time detection transformer RT-DETR in terms of mAP50, mAP50-95, and precision, while also exhibiting lower GFLOPs (Global Floating-Point Operations). This indicates that the method effectively controls the computational load while maintaining or even improving detection accuracy, achieving a better balance between detection performance and computational complexity. Therefore, the method of this invention has significant application value in scenarios where both detection accuracy and deployment efficiency are critical.

[0079] From the perspectives of method design and experimental results, the main advantage of this invention compared to existing supervised and semi-supervised detection methods lies in proposing a target-level discriminative prior reliability transfer mechanism for remote sensing small target detection. This invention employs Grounding DINO as the teacher model and the Real-Time Detection Transformer (RT-DETR) as the student model. Instead of directly using the teacher model's output on unlabeled images as pseudo-labels, it uniformly models them as candidate target-level discriminative priors containing bounding boxes, class probability distributions, and targetability scores. Based on this, this invention performs joint reliability judgment on candidate priors from two dimensions: class determinism and localization stability. Only semantically clear and geometrically stable candidate targets are retained. Furthermore, the targetability scores output by the teacher are combined to assign sample-level training weights to reliable priors, thereby achieving hierarchical screening and differentiated utilization of unlabeled sample supervision information and reducing the adverse effects of noise supervision propagation on student model training.

[0080] Experimental results show that the proposed method achieves a mAP of 76.9% on the DIOR remote sensing target detection dataset, outperforming various existing supervised detection methods and semi-supervised detection baselines, thus verifying the effectiveness of the proposed method in remote sensing target detection tasks under low-label conditions. Further ablation experiments demonstrate that category determinism, localization stability, and target-aware weighting strategies all contribute significantly to performance improvements, with localization stability showing the most significant improvement in mAP75. This indicates that the proposed method not only improves overall detection accuracy but also effectively enhances target localization quality. This is because small remote sensing targets typically exhibit characteristics such as small scale, blurred boundaries, complex backgrounds, and volatile spatial responses, making it difficult to guarantee pseudo-supervision quality solely based on the classification confidence output of the teacher model; however, this invention introduces cross-view... Figure 1 Consistent localization stability constraints are used to filter out candidate targets with unstable geometric responses, making the unsupervised supervision signals received by the student model more reliable in spatial location. Therefore, more significant performance gains are achieved under strict localization evaluation metrics.

[0081] Furthermore, this invention achieves an effective improvement in the utilization of unlabeled samples without altering the main structure, parameter scale, or computational complexity of the student detector. Experimental results show that the performance improvement of the method in this invention mainly stems from the reliable prior screening and weighted transfer mechanism itself, rather than from additional model structure expansion or the introduction of complex computational modules. Therefore, this invention can improve the accuracy, stability, and interpretability of semi-supervised remote sensing small target detection while maintaining consistency with the basic framework of RT-DETR, and has good practical application value.

[0082] While specific embodiments of the present invention have been described above, those skilled in the art should understand that these are merely illustrative examples. Various changes or modifications can be made to these embodiments without departing from the principles and essence of the present invention. Therefore, the scope of protection of the present invention is defined by the appended claims.

[0083] The above embodiments are preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Various modifications or variations that can be made by those skilled in the art without creative effort within the scope of the appended claims are still within the scope of protection of the present invention.

Claims

1. A semi-supervised remote sensing small target detection method based on target level discriminative prior transfer, characterized in that: First, a teacher model is used to generate candidate target-level discriminative priors on unlabeled images; Secondly, based on the candidate target-level discrimination prior, joint reliability judgment is performed from two dimensions: category certainty and positioning stability. Sample-level training weights are assigned based on the target score to generate a reliable prior set. Finally, the student model performs supervised training and unsupervised prior transfer training according to the input sample type and in combination with the reliable prior set. The trained student model is then used to detect targets in the remote sensing image to be inspected. For any unlabeled image The candidate target-level discriminative prior set output by the teacher model Represented as in, Indicates the number of candidate targets. Indicates the first Predicted bounding boxes of candidate targets, This represents the corresponding category probability distribution. This represents the corresponding goal score, indicating the confidence level that the candidate goal belongs to the prospective goal. Indicates the total number of categories; When determining joint reliability, firstly, a category determinism score is calculated based on the category probability distribution, thereby generating a category availability mask. Simultaneously, a location stability score is calculated based on the predicted bounding box, thereby generating a location availability mask. Finally, a joint availability mask is generated based on the category availability mask and the location availability mask. Then, based on the target score, sample-level training weights are generated. Then the image Corresponding reliable prior set Represented as The reliable prior set corresponding to the entire unlabeled sample set Represented as in, Indicates the number of unlabeled samples. This indicates the number of candidate targets.

2. The semi-supervised remote sensing small target detection method based on target-level discriminative prior transfer as described in claim 1, characterized in that: Based on category deterministic threshold Define category availability mask for in, The information entropy function is represented by... Indicates the category certainty score; Based on positioning stability threshold Define the location availability mask for in, This represents the positioning stability score. Indicates the intersection-union ratio of the bounding boxes. Indicates unlabeled images Apply geometric transformation The augmented image is then obtained, and the teacher model's predicted bounding box is located on the augmented image. Representing geometric transformations inverse transform; Calculate the joint availability mask using the following formula. , in, Indicates the first Whether the candidate target is retained.

3. The semi-supervised remote sensing small target detection method based on target-level discriminative prior transfer as described in claim 2, characterized in that: Calculate the sample-level training weights using the following formula. , in, Indicates the first The sample-level training weights corresponding to each candidate target.

4. The semi-supervised remote sensing small target detection method based on target-level discriminative prior transfer as described in claim 1, characterized in that: When training the student model, supervised learning is performed directly based on the real annotations for labeled samples; For unlabeled samples, the parameters corresponding to the candidate targets are first generated using the student model, and then unsupervised prior transfer training is performed in combination with a reliable prior set.

5. The semi-supervised remote sensing small target detection method based on target-level discriminative prior transfer as described in claim 4, characterized in that: For labeled samples The student model is directly based on real annotations. Supervised learning is performed, and its loss function is... Defined as in, This indicates supervised classification loss. This indicates the supervised bounding box regression loss. Indicates supervised regression loss The weighting coefficients.

6. The semi-supervised remote sensing small target detection method based on target-level discriminative prior transfer as described in claim 4, characterized in that: For unlabeled samples The student model output is represented as in, , and These represent the student model in the first... Predicted bounding boxes, class probability distributions, and target scores for each candidate target. Indicates the number of candidate targets; First, for each sample from the reliable prior set, a one-to-one matching strategy based on maximizing the bounding box IoU is adopted to find the sample in the unlabeled sample set that is closest to the geometric position of the candidate target, thereby achieving sample matching. These samples are then used together for training the student model. Each candidate target is matched with at most one sample from the reliable prior set. Secondly, the unsupervised migration loss is calculated using the following formula. ; in, Indicates the Kullback-Leibler divergence. and These represent the predicted bounding boxes output by the student model and the teacher model, respectively. This represents the target score output by the teacher model. This represents the target score output by the student model. Indicates the weight of location migration loss. Indicates the weight of the target migration loss. , , These represent the category migration loss, location migration loss, and destination migration loss, respectively. Represents the joint availability mask. This indicates the generation of sample-level training weights.

7. The semi-supervised remote sensing small target detection method based on target-level discriminative prior transfer as described in claim 1, characterized in that: The teacher model is set to Grounding DINO, and the student model is set to RT-DETR.

Images