Shielding target high-precision identification method based on data enhancement
By using generative adversarial network data augmentation and an improved bidirectional feature pyramid network, combined with a composite loss function, the problems of semantic information loss and feature confusion in occluded target detection are solved, achieving high-precision occluded target recognition.
Patent Information
- Application Number
- CN202610005824.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-01-05
- Publication Date
- 2026-04-17
AI Technical Summary
Existing technologies struggle to effectively express the complete semantic information of occluded targets in complex occlusion environments. They lack modeling of the unique spatial context relationships of occluded targets and are prone to feature confusion and boundary blurring in multi-target occlusion situations, affecting detection accuracy and localization accuracy.
A data augmentation method based on generative adversarial networks is used to generate a training dataset for the occluded target detection task. An improved bidirectional feature pyramid network and an adaptive spatial feature fusion mechanism are constructed. Multi-task joint training is carried out in combination with a composite loss function, and high-precision recognition results are output through three-level post-processing filtering.
It significantly improves detection accuracy under different types and degrees of occlusion, increasing vehicle recognition accuracy by 15.2% and pedestrian detection accuracy by 18.3% in densely populated scenes, and effectively solves the problems of feature confusion and boundary blurring under multi-target occlusion.
Smart Images

Figure CN121884059A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the fields of computer vision and deep learning technology, and in particular relates to a high-precision method for identifying occluded targets based on data augmentation. Background Technology
[0002] With the rapid development of computer vision and deep learning technologies, object detection technology has been widely used in fields such as autonomous driving, intelligent security, and robot navigation. However, in real-world complex environments, target objects are often partially or completely occluded by other objects, resulting in the loss of key feature information of the target. This seriously affects the accuracy and robustness of detection algorithms, becoming a key technical bottleneck restricting the practical application of object detection technology.
[0003] To address the problem of occluded target detection, Chinese patent application CN120047782A discloses an occluded target detection method based on a feature pyramid network. This method extracts features from a preprocessed image using a backbone network to obtain multi-scale shallow features. Then, a weighted feature pyramid network is used to fuse these multi-scale shallow features to obtain multi-scale deep features that eliminate the imbalance between semantic and positional information. Finally, the target category and location information of each prediction box are determined based on the multi-scale deep features, and the optimal detection result is output by removing redundant detection results. This method, through the design of the weighted feature pyramid network, improves the problem of the imbalance between semantic and positional information in traditional feature pyramid networks to a certain extent, thereby increasing the detection accuracy of occluded targets.
[0004] While the aforementioned existing technologies have addressed some of the imbalance between semantic and positional information through weighted feature pyramid networks, they still suffer from the following technical shortcomings: First, these methods primarily rely on traditional backbone networks for feature extraction. In cases of severe occlusion, the extracted features have limited representational capabilities due to significant loss of target information, making it difficult to fully express the complete semantic information of the occluded target. Second, although weighted feature pyramid networks consider the balance between semantic and positional information, they lack a dedicated modeling mechanism for the spatial contextual relationships and occlusion patterns specific to occluded targets. Furthermore, when dealing with complex occlusion scenarios, especially when multiple targets occlude each other, these methods are prone to feature confusion and boundary blurring, affecting the final detection accuracy and localization accuracy. Summary of the Invention
[0005] The purpose of this invention is to provide a high-precision identification method for occluded targets based on data augmentation, which can maintain stable high-precision performance under different types and degrees of occlusion.
[0006] To achieve the objective of this invention, a high-precision method for identifying occluded targets based on data augmentation is provided, comprising the following steps:
[0007] S1. Data Acquisition and Preprocessing: Standardize the original image dataset and the image to be detected to obtain a standardized dataset of consistent quality.
[0008] S2. Data augmentation based on generative adversarial networks: Generative adversarial network data augmentation method is applied to the standardized dataset to generate an augmented training dataset for the occluded target detection task;
[0009] S3. Construct an occluded target detection network: Construct a detection network model for the occluded target detection task, including a feature extraction layer, a bidirectional feature fusion layer, and a prediction output layer. The feature information of different resolution levels is integrated through multi-scale feature extraction and bidirectional feature pyramid fusion.
[0010] S4. Multi-task joint training: Based on the enhanced training dataset, the detection network model is trained using a composite loss function joint training strategy and a progressive multi-scale training strategy; after training, the image to be detected is input into the optimized detection network model, and the detection result is output.
[0011] S5. Accurate identification of occluded targets: The detection results are subjected to three levels of post-processing filtering. First, redundant detection boxes are removed by non-maximum suppression algorithm. Then, confidence threshold filtering is performed to eliminate low-quality predictions. Finally, the category determination is completed, and the category label, confidence score and bounding box coordinates of each target are output.
[0012] A non-transitory computer-readable storage medium storing computer instructions for causing the computer to perform the above-described data-enhanced high-precision identification method for occluded targets.
[0013] A computer program product includes computer program instructions that, when executed on a computer, cause the computer to perform the aforementioned data-enhanced high-precision identification method for occluded targets.
[0014] The significant advancement of this invention compared to existing technologies lies in:
[0015] (1) In view of the shortcomings of traditional backbone networks in expressing the complete semantic information of occluded targets under severe occlusion, this invention effectively enhances the completeness of the feature expression of occluded targets and improves the ability to compensate for missing information by constructing an improved bidirectional feature pyramid network and an adaptive spatial feature fusion mechanism.
[0016] (2) In view of the shortcomings of existing weight feature pyramid networks in lacking specific modeling of the spatial context relationship and occlusion pattern of occluded targets, this invention establishes a data augmentation method for multi-scene fusion based on generative adversarial network, generates style code according to scene label, realizes balanced data generation for different occlusion scenes, and provides rich occlusion pattern learning samples for the model.
[0017] (3) In view of the technical problem that feature confusion and boundary blurring are easy to occur when multiple targets occlude each other, the present invention designs a composite loss function including an improved attraction-repulsion mechanism, which effectively solves the problems of inaccurate prediction box localization, false detection and missed detection in dense occlusion environment;
[0018] (4) Through verification, the recognition accuracy of the present invention under various occlusion scenarios is significantly improved compared with the existing technology; the accuracy rate in vehicle recognition application reaches 89.7%, which is 15.2% higher than the existing method; the pedestrian detection accuracy in dense crowd scenarios is improved by 18.3%; therefore, the present method can maintain stable high-precision performance under different occlusion types and degrees.
[0019] (5) The present invention expands the original training set from 5,000 images to 25,000 balanced scene data through the data augmentation method of generative adversarial networks, and reduces the sample distribution variance of various occlusion scenes from 0.42 in the prior art to 0.08.
[0020] To more clearly illustrate the functional characteristics and structural parameters of the present invention, further explanation is provided below in conjunction with the accompanying drawings and specific embodiments. Attached Figure Description
[0021] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:
[0022] Figure 1 This is a flowchart of the steps of the present invention;
[0023] Figure 2 This is a diagram of the occlusion target detection network architecture of the present invention;
[0024] Figure 3 (a)-(d) are internal structure diagrams of the Bidirectional Feature Pyramid Network (BiFPN) and the Adaptive Spatial Feature Fusion Mechanism (ASFF) of this invention. Detailed Implementation
[0025] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0026] This invention provides a high-precision method for identifying occluded targets based on data augmentation, combined with... Figure 1 This includes the following steps:
[0027] S1. Data Acquisition and Preprocessing: Standardize the original image dataset and the image to be detected to obtain a standardized dataset of consistent quality. This preprocessing eliminates the differences in resolution, illumination and contrast of the original images, providing a standardized dataset of consistent quality for subsequent data augmentation and model training.
[0028] S2. Data Augmentation Based on Generative Adversarial Networks: Generative adversarial network data augmentation methods are applied to the standardized dataset to generate an augmented training dataset for the occluded target detection task; this dataset covers scenes with different degrees and types of occlusion; the augmented dataset is used for network training in step S4 to solve the problem of uneven distribution of occluded scenes in the original training samples.
[0029] S3. Construct an occluded target detection network: A three-layer detection network model is constructed for the occluded target detection task, including a feature extraction layer, a bidirectional feature fusion layer, and a prediction output layer. Feature information at different resolution levels is integrated through multi-scale feature extraction and bidirectional feature pyramid fusion. This network architecture improves the feature representation ability and bounding box localization accuracy of partially occluded targets by fusing high-level semantic information and low-level positional details.
[0030] S4. Multi-task joint training: Based on the enhanced training dataset, the detection network model is trained using a composite loss function joint training strategy and a progressive multi-scale training strategy; after training, the image to be detected is input into the optimized detection network model, and the detection result is output.
[0031] S5. Accurate Identification of Occluded Targets: The detection results undergo three levels of post-processing filtering. First, redundant detection boxes are removed using a non-maximum suppression algorithm. Then, confidence threshold filtering is performed to eliminate low-quality predictions. Finally, category determination is completed, outputting the category label, confidence score, and bounding box coordinates for each target. This step, through progressive filtering of redundancy suppression, confidence screening, and category determination, ensures the accuracy and reliability of the final identification results.
[0032] The standardization preprocessing of S1 includes accurate annotation (annotation based on occlusion type, occlusion degree and scene parameters), size normalization (640×640 resolution), color averaging (eliminating the influence of lighting) and contrast enhancement.
[0033] The generative adversarial network data augmentation method of S2 generates occluded target images through an encoder-decoder structure generator, evaluates the realism of the images through a multi-scale discriminator, and performs adversarial training.
[0034] Combination Figure 2 The feature extraction layer, feature fusion layer, and prediction output layer of S3 are specifically as follows: The feature extraction layer extracts multi-scale feature maps, including an input module, a multi-level feature extraction module, and a feature enhancement module connected in sequence; the input module includes convolutional layers and a C2f module (an efficient feature extraction module introduced in the YOLOv8 object detection network) to perform initial feature extraction on the input image; the multi-level feature extraction module includes multiple cascaded convolutional layers and a C2f module, which extracts features at different scales through progressive downsampling, and outputs three feature maps P3, P4, and P5 with different resolutions; the feature enhancement module adds a Spatial Pyramid Pooling Fast (SPPF) module after the P5 feature map, which enhances the feature extraction capability of small targets through multi-scale pooling operations, and embeds an attention mechanism module for occluded regions during each level of feature extraction to enhance the feature response of occluded targets;
[0035] Combination Figure 3 (a)-(d), the feature fusion layer adopts a bidirectional feature pyramid fusion network to realize bidirectional information transmission from top to bottom and from bottom to top, and dynamically adjusts the fusion weights of features at different scales through an adaptive spatial feature fusion mechanism;
[0036] The prediction output layer uses the prediction head network to output the location coordinates, class probability, and confidence level of each detected target.
[0037] The input to the joint training strategy of the composite loss function in S4 is the predicted bounding box coordinates, class probability, and confidence score output by the detection network model, as well as the real annotation information in the augmented training dataset. The output is a composite loss value that combines localization loss, classification loss, and confidence loss, which is used for backpropagation to update network parameters. The input to the progressive multi-scale training strategy is the training images in the augmented training dataset, and the output is the optimized network model parameters after multi-scale training.
[0038] The progressive multi-scale training strategy employs small-scale images (320×320 resolution) for rapid convergence in the initial stage, gradually increases image resolution (480×480, 640×640 resolution) in the intermediate stage to enhance detail learning ability, and uses multi-scale hybrid training (randomly sampling resolutions from 320×320 to 640×640) to enhance model generalization in the final stage. The composite loss function joint training strategy works in conjunction with the progressive multi-scale training strategy. Specifically, in each stage of the progressive multi-scale training, a composite loss function is used to calculate the training loss for the current stage. The network undergoes a coarse-to-fine learning process, with parameter updates performed concurrently. The initial stage uses small-scale images and a composite loss function to enable the network to quickly learn the basic feature representations and localization capabilities of the target. The intermediate stage, building upon the convergence of the initial stage, gradually increases the image resolution, allowing the network to further learn the detailed features and precise boundaries of the target. The final stage involves multi-scale hybrid training, enabling the network to adapt to different scale inputs and achieve generalization performance. These three stages progress sequentially, with the training results of the previous stage serving as the initial parameters for the next. Through this progressive learning process, a robust network model capable of detecting targets with varying degrees of occlusion is ultimately obtained.
[0039] The joint training strategy for the composite loss function of S4 is as follows:
[0040] Construct a composite loss function:
[0041] ;
[0042] in, To improve positioning loss; For classification loss; For confidence loss; , , The target weight coefficients of the evaluation detection boxes are set separately; the target weight coefficients of the evaluation detection boxes , , The default settings are 1.0, 1.0, and 1.0.
[0043] The improved positioning loss Including attractions and exclusion term The attraction-repulsion mechanism causes positive bounding boxes to move closer to the ground truth bounding boxes, while negative bounding boxes move away from them.
[0044] ;
[0045] ;
[0046] in, Here, is the index of the positive sample detection box, and j is the index of the positive sample detection box. For the first The Euclidean distance between a positive sample detection box and its corresponding ground truth box. Let be the Euclidean distance between the j-th negative sample detection box and the nearest ground truth box. The distance threshold of the attraction term, The distance boundary threshold for the exclusion term;
[0047] The attraction term causes the positive sample detection boxes that match the real target to gradually converge towards the real box position, improving the bounding box localization accuracy and reducing false detections; the repulsion term causes the negative sample detection boxes that do not match the real target to move away from the real target area, suppressing false detections in the background area; this attraction-repulsion mechanism effectively distinguishes the boundaries of adjacent targets in densely occluded scenes, reducing false detections and false detections caused by target overlap.
[0048] The classification loss adopts the focus loss function. By introducing a modulation factor, the loss weight of easily classified samples is reduced and the loss weight of difficult-to-classify samples is increased. This makes the model training focus on samples that are difficult to classify, such as those with severe occlusion and unclear features, thus solving the problem of imbalance between positive and negative samples.
[0049] The S5 multi-level post-processing filtering includes a three-level processing flow: The first level processes the bounding box coordinates using the non-maximum suppression (NMS) algorithm to remove redundant detection boxes with high overlap; the IoU threshold was experimentally determined to be 0.5. The second level filters the confidence scores of the remaining detection boxes, eliminating low-quality predictions below the confidence threshold; the confidence threshold was experimentally determined to be 0.7. The third level determines the category label of the detection results that have passed the first two levels of filtering to determine the final category of each target. After three levels of filtering, the category label, confidence score, and bounding box coordinates of each target are output.
[0050] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.
[0051] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A high-precision method for identifying occluded targets based on data augmentation, characterized in that, Includes the following steps: S1. Data Acquisition and Preprocessing: Standardize the original image dataset and the image to be detected to obtain a standardized dataset of consistent quality. S2. Data augmentation based on generative adversarial networks: Generative adversarial network data augmentation method is applied to the standardized dataset to generate an augmented training dataset for the occluded target detection task; S3. Construct an occluded target detection network: Construct a detection network model for the occluded target detection task, including a feature extraction layer, a bidirectional feature fusion layer, and a prediction output layer. The feature information of different resolution levels is integrated through multi-scale feature extraction and bidirectional feature pyramid fusion. S4. Multi-task joint training: Based on the enhanced training dataset, the detection network model is trained using a composite loss function joint training strategy and a progressive multi-scale training strategy. After training, the image to be detected is input into the optimized detection network model, and the detection result is output. S5. Accurate identification of occluded targets: The detection results are subjected to three levels of post-processing filtering. First, redundant detection boxes are removed by non-maximum suppression algorithm. Then, confidence threshold filtering is performed to eliminate low-quality predictions. Finally, the category is determined and the category label, confidence score and bounding box coordinates of each target are output.
2. The high-precision identification method for occluded targets based on data augmentation according to claim 1, characterized in that, The standardized preprocessing of S1 includes labeling based on occlusion type, occlusion degree and scene parameters, size normalization, color averaging and contrast enhancement.
3. The high-precision identification method for occluded targets based on data augmentation according to claim 1, characterized in that, The generative adversarial network data augmentation method of S2 generates occluded target images through an encoder-decoder structure generator, evaluates the realism of the images through a multi-scale discriminator, and performs adversarial training.
4. The high-precision identification method for occluded targets based on data augmentation according to claim 1, characterized in that, The feature extraction layer, feature fusion layer, and prediction output layer of S3 are as follows: The feature extraction layer extracts multi-scale feature maps, including an input module, a multi-level feature extraction module, and a feature enhancement module connected in sequence; the input module includes convolutional layers and a C2f module to perform initial feature extraction on the input image; the multi-level feature extraction module includes multiple cascaded convolutional layers and a C2f module, which extract features at different scales through progressive downsampling, and output three feature maps P3, P4, and P5 with different resolutions; the feature enhancement module adds a spatial pyramid pooling module after the P5 feature map, which enhances the feature extraction capability of small targets through multi-scale pooling operations, and embeds an attention mechanism module for occluded regions during each level of feature extraction to enhance the feature response of occluded targets; the feature fusion layer uses a bidirectional feature pyramid fusion network to achieve bidirectional information transmission from top to bottom and bottom to top, and dynamically adjusts the fusion weights of features at different scales through an adaptive spatial feature fusion mechanism; the prediction output layer uses a prediction head network to output the location coordinates, class probability, and confidence score of each detected target.
5. The high-precision identification method for occluded targets based on data augmentation according to claim 1, characterized in that, The input to the joint training strategy of the composite loss function in S4 is the predicted bounding box coordinates, class probability, and confidence score output by the detection network model, as well as the real annotation information in the augmented training dataset. The output is a composite loss value that combines localization loss, classification loss, and confidence loss, which is used for backpropagation to update network parameters. The input to the progressive multi-scale training strategy is the training images in the augmented training dataset, and the output is the optimized network model parameters after multi-scale training. The progressive multi-scale training strategy employs a three-stage approach: initial stage using small-scale images for rapid convergence; intermediate stage gradually increasing image resolution to enhance detail learning capabilities; and final stage using multi-scale hybrid training to improve model generalization. The composite loss function joint training strategy works in conjunction with the progressive multi-scale training strategy. Specifically, in each stage of progressive multi-scale training, a composite loss function is used to calculate the training loss and update parameters. Initially, small-scale images combined with the composite loss function enable the network to quickly learn the basic feature representation and localization capabilities of the target. In the intermediate stage, based on the convergence achieved in the initial stage, image resolution is gradually increased to allow the network to further learn the detailed features and precise boundaries of the target. The final stage uses multi-scale hybrid training to adapt the network to different scale inputs and achieve generalization performance. These three stages progress sequentially, with the training results of the previous stage serving as the initialization parameters for the next stage. Through this coarse-to-fine progressive learning process, an optimized network model with robust detection capabilities for targets with varying degrees of occlusion is ultimately obtained.
6. The high-precision identification method for occluded targets based on data augmentation according to claim 1, characterized in that, The joint training strategy for the composite loss function of S4 is as follows: Construct a composite loss function: ; in, To improve positioning loss; For classification loss; For confidence loss; , , The target weight coefficients are set separately for the evaluation detection boxes; The improved positioning loss Including attractions and exclusion term The attraction-repulsion mechanism causes positive bounding boxes to move closer to the ground truth bounding boxes, while negative bounding boxes move away from them. ; ; in, Here, is the index of the positive sample detection box, and j is the index of the positive sample detection box. For the first The Euclidean distance between a positive sample detection box and its corresponding ground truth box. Let be the Euclidean distance between the j-th negative sample detection box and the nearest ground truth box. The distance threshold of the attraction term, This is the distance boundary threshold for the exclusion term.
7. The high-precision identification method for occluded targets based on data augmentation according to claim 6, characterized in that, The classification loss uses a focus loss function, which reduces the loss weight of easily classified samples and increases the loss weight of difficult-to-classify samples by introducing a modulation factor.
8. The high-precision identification method for occluded targets based on data augmentation according to claim 1, characterized in that, The multi-level post-processing filtering of S5 includes a three-level processing flow: the first level processes the bounding box coordinates using a non-maximum suppression algorithm to remove redundant detection boxes with high overlap; the second level filters the confidence scores of the remaining detection boxes to eliminate low-quality predictions below the confidence threshold. The third level determines the category label of the detection results that have passed the first two levels of screening, and determines the final category of each target; After three levels of filtering, the output includes the category label, confidence score, and bounding box coordinates for each target.
9. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium stores computer instructions for causing the computer to perform the method according to any one of claims 1 to 8.
10. A computer program product comprising computer program instructions, characterized in that, When the computer program instructions are executed on a computer, the computer causes the computer to perform the method as described in any one of claims 1 to 8.
Citation Information
Patent Citations
Shielding target detection method based on feature pyramid network
CN120047782A
