Density center guide perception enhancement method for detection of dense small targets in unmanned aerial vehicle image

Abstract

The invention discloses a density center guided perception enhancement method for detection of dense small targets in an unmanned aerial vehicle image, and the method comprises the steps: obtaining an aerial image of an unmanned aerial vehicle, inputting the aerial image of the unmanned aerial vehicle into a preset backbone network and an encoder, and carrying out the extraction to obtain a multi-scale feature map; selecting a specified layer feature map from the multi-scale feature map, and inputting the specified layer feature map into a density guide target center heat map generator; density features are generated through number estimation branches, and a target center thermodynamic diagram is generated through the density features and used for supervising generation of target center features; inputting the density feature and the target center feature into a density-center feature enhancement module, and processing the density feature and the target center feature by the density-center feature enhancement module to obtain an enhanced feature; fusing the enhanced features with the multi-scale feature map to obtain a fused feature set; and inputting the fusion feature set into a preset decoder and a prediction head, and outputting a detection result of the target in the unmanned aerial vehicle image after the fusion feature set is processed by the decoder and the prediction head.

Description

Technical Field

[0001] This invention belongs to the field of UAV image processing and target detection technology, and particularly relates to a density center-guided perception enhancement method for detecting dense small targets in UAV images. Background Technology

[0002] Drone aerial imagery, with its wide field of view and high coverage, has been widely used in various fields such as public safety monitoring, agricultural growth monitoring, and disaster emergency response. However, targets in aerial images are often small in size, densely packed, and easily obstructed, posing significant challenges to target detection tasks. To address these challenges, researchers have conducted extensive studies on improving the accuracy and efficiency of detecting dense small targets, resulting in three main technical directions: First, feature enhancement methods, such as HRDNet which preserves spatial information by fusing multi-resolution inputs and multi-scale features, SAIC which uses GAN super-resolution technology to reduce the difficulty of detecting low-resolution targets, and SPAR which combines visual-linguistic semantic cues to improve the detection effect in complex scenes. These methods focus on optimizing the quality of global features to enhance the perception of small targets. Second, real-time detection optimization methods, such as RemDet which builds an efficient detection architecture based on minimizing information loss, and QueryDet which balances speed and accuracy by first predicting the coarse location of small targets on low-resolution features and then accurately detecting them on high-resolution features, aiming to adapt to the limited computing resources on UAVs. Third, region cropping and integration methods, such as DREN which adaptively crops difficult regions of images, ClusDet which crops based on target clustering regions, DMNet which uses density maps to guide region segmentation, AdaZoom which generates focused regions through reinforcement learning, and UFPMP-Det and Zoom&Reasoning. Det stitches the cropped regions into a single image to achieve unified detection. This type of method improves the utilization of small target features by focusing on local regions.

[0003] While existing technologies have made some progress in the field of UAV image target detection, they still have significant shortcomings and cannot fully meet the needs of detecting dense small targets. Feature enhancement methods often focus on global feature modeling, neglecting the explicit characterization of dense target regions. In scenes with highly crowded targets, they cannot effectively distinguish the feature boundaries of adjacent small targets, easily leading to missed detections or misjudgments. Real-time detection optimization methods often sacrifice fine-grained feature quality to improve efficiency, making it difficult to accurately capture detailed information of dense small targets, thus limiting detection reliability. Region cropping and integration methods rely on multi-stage processing, which not only increases inference time costs but also loses some global contextual information during cropping and stitching, affecting the model's perception of the overall target distribution and reducing detection robustness. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention proposes a density center-guided perception enhancement method for detecting dense small targets in UAV imagery. This method can simultaneously achieve global target distribution modeling and local instance detail representation within a single detection framework, balancing detection accuracy, efficiency, and robustness, thereby solving the problems existing in the prior art.

[0005] To achieve the above objectives, this invention provides a density center-guided perception enhancement method for detecting dense small targets in UAV imagery, comprising: S1: Acquire drone aerial images, input the drone aerial images into a preset backbone network and encoder, and extract multi-scale feature maps; S2: Select a specified layer feature map from the multi-scale feature map, input the specified layer feature map into the density-guided target center heatmap generator, and generate density features through the quantity estimation branch. The density-guided target center heatmap generator combines the local target density information contained in the density features to generate a density-adaptive target center heatmap for supervising the generation of target center features. S3: Input the density feature and the target center feature into the density-center feature enhancement module. The density-center feature enhancement module first performs dilated convolution processing on the density feature to expand the receptive field, and performs spatial attention weight generation processing on the target center feature to highlight the target center region. Then, it calculates the similarity between the processed density feature and the processed target center feature to obtain a fusion weight map. Finally, based on the fusion weight map, it performs weighted fusion of the two types of processed features and performs channel weighting to obtain the enhanced feature. S4: Fuse the enhanced features with the multi-scale feature map to obtain a fused feature set; S5: Input the fused feature set into a preset decoder and prediction head, and output the detection results of dense small targets in the UAV image after processing by the decoder and prediction head.

[0006] Preferably, in step S1, the drone aerial image is input into a preset backbone network and encoder to extract multi-scale feature maps, specifically including: The drone aerial images are input into the backbone network, and initial features at different resolutions are obtained after feature extraction layer by layer by the backbone network. Each of the initial features is input into the encoder. After feature encoding and scale alignment by the encoder, the multi-scale feature map containing feature information at different scales is output.

[0007] Preferably, step S2, selecting a specified layer feature map from the multi-scale feature map, specifically includes: The spatial detail retention of each layer of the multi-scale feature map is analyzed, and the layer feature map with the most complete spatial detail retention and the ability to reflect the local information of the small target is selected as the specified layer feature map.

[0008] Preferably, the density-guided target center heatmap generator in step S2 combines the local target density information contained in the density features to generate a density-adaptive target center heatmap, specifically including: The density-guided target center heatmap generator calculates the local target density within each real target bounding box based on the density features. The variance parameter of the Gaussian kernel is adjusted according to the local target density to generate a Gaussian response at the center of the real target bounding box. A coefficient related to the target area is introduced to enhance the response of small targets in the heat map, and finally the density-adaptive target center heat map is generated iteratively.

[0009] Preferably, step S3 involves calculating the similarity between the processed density features and the processed target center features to obtain a fused weight map, specifically including: Convolution operations are performed on the processed density features and the processed target center features respectively to keep the number of channels of the two types of features consistent. For two types of features with the same number of channels, multiply them element by element and sum the results along the channel dimension; The summation result is input into the activation function for normalization to obtain the fusion weight map that reflects the similarity between the two types of features.

[0010] Preferably, step S3 involves weighted fusion of the two processed features based on the fusion weight map, specifically including: The fusion weight map is multiplied element-wise with the processed density features to obtain the first fusion part dominated by the density features; Calculate the difference between 1 and the fusion weight map, and multiply the difference element-wise with the processed target center feature to obtain the second fusion part dominated by the target center feature; The first fusion portion and the second fusion portion are added element by element to obtain a fusion feature that takes into account both regional density information and target center information.

[0011] Preferably, step S3 involves channel weighting of the weighted fusion result, specifically including: The weighted fusion results are subjected to feature statistics at the channel dimension, and the contribution of each channel feature to target detection is calculated. Assign weight coefficients to the corresponding channels based on their respective contributions; Each channel feature is multiplied element-wise with its corresponding weight coefficient to obtain the enhanced features that strengthen key features and suppress redundant features.

[0012] Preferably, in step S4, the enhanced features are fused with the multi-scale feature map to obtain a fused feature set, specifically including: The enhanced features are scaled so that the resolution of the enhanced features is consistent with the resolution of the corresponding layer feature map in the multi-scale feature map; The enhanced features with consistent resolution and the corresponding layer feature maps are processed by spatial attention and channel attention mechanisms to obtain fused features of each layer that integrate density-center enhancement information. The fusion features of all layers are summarized to form the fusion feature set.

[0013] Preferably, step S5, which outputs the target detection result in the UAV image after processing by the decoder and prediction head, specifically includes: The fused feature set is input into the decoder, which performs multi-stage decoding on the fused feature set in conjunction with a preset dynamic query, and outputs intermediate predicted features at each stage to optimize the decoding process. The decoding features from the final stage are input into the prediction head, which performs target category prediction and bounding box regression on the decoding features from the final stage, and outputs the detection result containing target category and location information.

[0014] Preferably, the method further includes a model training step: Obtain a training dataset, which contains multiple drone aerial sample images labeled with target categories and bounding boxes; Based on the training dataset, and with a preset total loss function as the optimization objective, the overall model consisting of the backbone network, encoder, density-guided target center heatmap generator, density-center feature enhancement module, decoder, and prediction head is iteratively trained. Continue until the model converges, resulting in a fully trained detection model; In this context, the backbone network, encoder, density-guided target center heatmap generator, density-center feature enhancement module, decoder, and prediction head mentioned in steps S1 to S5 are all corresponding components in the trained detection model.

[0015] Compared with the prior art, the present invention has the following advantages and technical effects: This invention uses a "density-guided target center heatmap generator that combines local target density information to generate a density-adaptive heatmap," which highlights target center information, reduces feature overlap and boundary blurring issues of adjacent dense targets, and reduces missed detections of overlapping targets. At the same time, the "density-center feature enhancement module integrates density features (global distribution) and target center features (local details)," so that feature representation takes into account both global target distribution and local instance details, enhancing the discrimination power for small, closely arranged targets. This solves the shortcomings of existing technologies that "ignore the modeling of dense regions for global features" and "sacrifice fine-grained features," significantly improving the detection accuracy of small targets.

[0016] The present invention eliminates the need for multi-stage region pruning and secondary detection, maintaining the end-to-end processing flow of "backbone network-encoder-decoder-prediction head" and avoiding additional inference time overhead. Furthermore, it preserves global contextual information through density features, compensating for the shortcomings of traditional pruning methods in "losing global information", enabling the model to still have stable detection capabilities in complex UAV scenarios (such as target occlusion and multi-scale mixing) and improving robustness.

[0017] This invention extracts multi-scale feature maps, enhances features, and fuses multi-scale features. It strengthens the feature response of small targets without weakening the feature representation of large targets. Even when small dense targets and large sparse targets coexist in UAV imagery, it can still achieve effective detection of all types of targets and has good scale adaptability. Attached Figure Description

[0018] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a schematic diagram of the overall architecture of a density center-guided perception enhancement method for detecting dense small targets in UAV imagery, according to an embodiment of the present invention. Figure 2 This is a schematic diagram comparing the effects of binary masks and Gaussian heatmaps in an embodiment of the present invention; Figure 3 This is a flowchart of the training steps in an embodiment of the present invention. Detailed Implementation

[0019] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0020] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0021] Example 1 This embodiment proposes DeCenter, a novel density-center-guided UAV target detection framework. DeCenter consists of two key modules: a Density-Guided Object Center Heatmap Generator (DOCHG) and a Density-Center Feature Enhancement module (DCFE), which together enhance the perception capability for small, dense targets. First, DOCHG adaptively generates a Gaussian kernel-based heatmap based on local density information, guiding the model to focus on the central neighborhood of targets within dense regions. This mechanism effectively reduces overlap between adjacent instances and alleviates the problem of missed detections under occlusion. Second, DCFE fuses complementary information from the density map and target centers, adaptively balancing regional target distribution with fine-grained localization capabilities. By fusing these signals, DCFE improves the quality of feature representation, making it not only more discriminative on dense, small targets but also suppressing background noise. Extensive experiments on two UAV benchmark datasets, VisDrone and UAVDT, validate the effectiveness of DeCenter. The proposed solution achieves significant improvements in both overall average accuracy and small target detection accuracy, reaching the performance level of existing state-of-the-art methods. These results demonstrate that density-center-guided feature enhancement provides an efficient and theoretically sound solution to the challenging problem of dense small target detection in UAV images.

[0022] The overall architecture of the method in this embodiment is as follows: Figure 1 As shown. Similar to the baseline method DQ-DETR, the framework employs a CNN backbone network, a deformable encoder-decoder, and a prediction head to achieve end-to-end target detection. Building upon this, this embodiment introduces two new components: (a) a density-guided target center heatmap generator (DOCHG, Figure 1 (a) This module uses density features generated by dynamic queries to generate heatmaps; (b) Density-Center Feature Enhancement Module (DCFE) Figure 1 (b) This module gradually shifts the focus of feature attention from a large area to specific target instances, thereby improving the model's ability to distinguish targets in dense scenes.

[0023] The overall process is as follows: After the input image is processed by a backbone network to extract multi-scale features, it is flattened and input into a deformable encoder. The encoded feature vectors are then reconstructed into multi-scale feature maps, where... Layer features are fed into DOCHG to generate a density-guided target center heatmap, which is used to supervise the generation of target center features. This target center feature is then fused with the density feature via DCFE to obtain an enhanced target-aware feature representation. Subsequently, these enhanced features are combined with the original multi-scale features and further input into the deformable decoder and prediction head. During decoding, the decoder works in conjunction with a corresponding number of dynamic queries to ultimately generate the detection result.

[0024] This embodiment provides a density center-guided perception enhancement method for detecting dense small targets in UAV imagery, specifically including: S1: Acquire drone aerial images, input the drone aerial images into a preset backbone network and encoder, and extract multi-scale feature maps; Further, in step S1, the drone aerial images are input into a preset backbone network and encoder to extract multi-scale feature maps, specifically including: The drone aerial images are input into the backbone network, and initial features at different resolutions are obtained after feature extraction layer by layer by the backbone network. Each of the initial features is input into the encoder. After feature encoding and scale alignment by the encoder, the multi-scale feature map containing feature information at different scales is output.

[0025] S2: Select a specified layer feature map from the multi-scale feature map, input the specified layer feature map into the density-guided target center heatmap generator, and generate density features through the quantity estimation branch. The density-guided target center heatmap generator combines the local target density information contained in the density features to generate a density-adaptive target center heatmap for supervising the generation of target center features. Further, step S2, selecting a specified layer feature map from the multi-scale feature map, specifically includes: The spatial detail retention of each layer of the multi-scale feature map is analyzed, and the layer feature map with the most complete spatial detail retention and the ability to reflect the local information of the small target is selected as the specified layer feature map.

[0026] Furthermore, the density-guided target center heatmap generator in step S2 combines the local target density information contained in the density features to generate a density-adaptive target center heatmap, specifically including: The density-guided target center heatmap generator calculates the local target density within each real target bounding box based on the density features. The variance parameter of the Gaussian kernel is adjusted according to the local target density to generate a Gaussian response at the center of the real target bounding box. A coefficient related to the target area is introduced to enhance the response of small targets in the heat map, and finally the density-adaptive target center heat map is generated iteratively.

[0027] Specifically, drone aerial images often contain dense regions where numerous targets overlap. In such cases, the feature boundaries extracted by the model become blurred, frequently leading to missed detections of overlapping instances. To address this issue—that blurred feature boundaries in dense regions may cause multiple targets to be identified as a single target—this embodiment proposes a density-guided target center heatmap. This heatmap highlights target center information, reduces feature ambiguity between overlapping targets, and facilitates accurate identification of densely distributed targets.

[0028] In remote sensing target detection research, some studies have used binary target masks to highlight target regions and suppress background noise. However, such binary masks are not suitable for UAV images because they often contain severe occlusion and significant overlap between targets, leading to missed detections even in dense scenes. To address this limitation, this embodiment replaces the binary mask with a Gaussian-based target heatmap. This heatmap can highlight target center information, generate clearer feature boundaries, and improve the recognition of overlapping targets.

[0029] A diagram comparing the effects of binary masks and Gaussian heatmaps is shown below. Figure 2 As shown, when targets appear densely, rectangular binary masks exhibit severe overlap. For example, when pedestrians walk side-by-side, their rectangular masks largely overlap, making it difficult to distinguish their individual positions. In contrast, Gaussian heatmaps mitigate this overlap through smooth local responses, thus more clearly indicating the location of each target. Density-guided target center heatmaps, by placing a Gaussian kernel at the center of each true bounding box, have a response value that gradually decays along the boundaries, as shown in the following formula: (1) in, Indicates the pixel position of the heatmap The value, Indicates the center coordinates of the bounding box. and Control the extent of the Gaussian distribution in the horizontal and vertical directions.

[0030] Since target overlap is more likely to occur in dense regions, this embodiment adaptively adjusts the Gaussian kernel parameters based on local target density. Specifically, a smaller variance is used in dense regions to generate sharper peaks, thereby better distinguishing adjacent targets; while a larger variance is used in sparse regions to provide a wider coverage. To quantify local density, this embodiment calculates the average density within each ground truth bounding box based on the density map: (2) in, and The density map represents the coordinates of the top-left and bottom-right corners of the bounding box. From density characteristics The density feature is obtained after channel mean normalization. Generated by a query quantity estimation branch, this branch estimates the number of targets and effectively encodes target density information. Based on average density... Gaussian variance and Defined as: (3) in and These are the width and height of the annotation box, respectively. The hyperparameters used to control Gaussian variance scaling.

[0031] To enhance the perception of small targets, this embodiment introduces an area-related coefficient in each target heatmap. This coefficient is defined as the logarithm of the ratio of the area of ​​the largest bounding box to the area of ​​the current bounding box. (4) in This represents the area of ​​the largest bounding box in the image. Taking the logarithm helps avoid this coefficient becoming too large when both extremely large and extremely small targets exist in the image. This is the enhanced heatmap. By integrating spatial density and target scale information, the model's ability to detect small, densely distributed targets is improved. Final heatmap. Get the current pixel by pixel and Maximum value iterative construction: (5) in Initialize to a zero matrix.

[0032] like Figure 1 As shown, DOCHG uses The layer feature map is used as input, and then processed through a series of dilated convolutions to obtain the predicted target center feature. And on After channel averaging and normalization, we get By minimizing Compared with the actual heat map The mean squared error (MSE) between samples is used for training to ensure accurate thermal features are generated during the inference phase. The loss function is defined as: (6) S3: Input the density feature and the target center feature into the density-center feature enhancement module. The density-center feature enhancement module first performs dilated convolution processing on the density feature to expand the receptive field, performs spatial attention weight generation processing on the target center thermal feature to highlight the target center region, then calculates the similarity between the processed density feature and the processed target center feature to obtain a fusion weight map, and finally performs weighted fusion of the two types of processed features based on the fusion weight map and performs channel weighting to obtain the enhanced feature. Further, step S3 calculates the similarity between the processed density features and the processed target center features to obtain the fused weight map, specifically including: Convolution operations are performed on the processed density features and the processed target center features respectively to keep the number of channels of the two types of features consistent. For two types of features with the same number of channels, multiply them element by element and sum the results along the channel dimension; The summation result is input into the activation function for normalization to obtain the fusion weight map that reflects the similarity between the two types of features.

[0033] Furthermore, step S3 involves weighted fusion of the two types of processed features based on the fusion weight map, specifically including: The fusion weight map is multiplied element-wise with the processed density features to obtain the first fusion part dominated by the density features; Calculate the difference between 1 and the fusion weight map, and multiply the difference element-wise with the processed target center feature to obtain the second fusion part dominated by the target center feature; The first fusion part and the second fusion part are added element by element to obtain a fusion feature that takes into account both regional density information and target center information.

[0034] Furthermore, step S3 involves channel weighting of the weighted fusion result, specifically including: The weighted fusion results are subjected to feature statistics at the channel dimension, and the contribution of each channel feature to target detection is calculated. Assign weight coefficients to the corresponding channels based on their respective contributions; Each channel feature is multiplied element-wise with its corresponding weight coefficient to obtain the enhanced features that strengthen key features and suppress redundant features.

[0035] Specifically, density maps can capture the spatial distribution of targets in an image, thereby quickly locating concentrated target regions. However, since density maps primarily focus on region-level information, they can only indicate the areas of interest and cannot depict fine-grained relationships between targets within a region or between targets and the background. To overcome this limitation, this embodiment proposes a density-center feature enhancement module. This module combines density features with density-guided target center features to enhance the target center features in the density representation, thereby improving the model's ability to perceive individual targets.

[0036] like Figure 1 As shown, DCFE first performs feature enhancement on both types of inputs separately. For density features... This embodiment uses dilated convolution to expand the receptive field and enrich the feature representation. For target center features... In this embodiment, spatial attention weights are generated to highlight the target center and suppress background noise and boundary information of overlapping targets, thereby making the target boundary clearer.

[0037] This embodiment calculates the enhanced density feature through element-wise multiplication. With enhanced central features The similarity between them is used to obtain the fusion weight graph: (7) in, This represents the sigmoid activation function. This represents the convolution operation. This represents the number of channels in the feature after the convolution operation.

[0038] Next, the two types of features are adaptively fused using this fusion weight map: (8) This design allows the model to adaptively maintain a balance between region density cues and target center cues. When a pixel is located in a dense region, the learned weights... Tendency to emphasize central thermal characteristics This allows for better differentiation of adjacent targets; however, in sparse regions, It tends to utilize density features To preserve contextual distribution information.

[0039] Compared to direct element-wise multiplication or simple concatenation, this complementary weighting mechanism more effectively mitigates information bias, suppresses noise, and ensures the joint preservation of the global distribution and fine-grained target boundaries. Finally, this embodiment performs channel weighting on the fused features to automatically learn the importance of different channels.

[0040] Compared to fusion methods based on direct multiplication, DCFE can better capture the relationship between density and target center, thereby achieving more effective feature integration and further improving detection performance.

[0041] S4: Fuse the enhanced features with the multi-scale feature map to obtain a fused feature set; Further, in step S4, the enhanced features are fused with the multi-scale feature map to obtain a fused feature set, specifically including: The enhanced features are scaled so that the resolution of the enhanced features is consistent with the resolution of the corresponding layer feature map in the multi-scale feature map; The enhanced features with consistent resolution and the corresponding layer feature maps are processed by spatial attention and channel attention mechanisms to obtain fused features of each layer that integrate density-center enhancement information. The fusion features of all layers are summarized to form the fusion feature set.

[0042] S5: Input the fused feature set into a preset decoder and prediction head, and output the detection result of the target in the UAV image after processing by the decoder and prediction head.

[0043] Further, in step S5, the detection results of the targets in the UAV image are output after processing by the decoder and prediction head, specifically including: The fused feature set is input into the decoder, which performs multi-stage decoding on the fused feature set in conjunction with a preset dynamic query, and outputs intermediate predicted features at each stage to optimize the decoding process. The final stage decoding features are input into the prediction head, which performs target category prediction and bounding box regression on the final stage decoding features, and outputs the detection result containing target category and location information.

[0044] Furthermore, the method also includes a model training step: Obtain a training dataset, which contains multiple drone aerial sample images labeled with target categories and bounding boxes; Based on the training dataset, and with a preset total loss function as the optimization objective, the overall model, consisting of the backbone network, encoder, density-guided target center heatmap generator, density-center feature enhancement module, decoder, and prediction head, is iteratively trained. Continue until the model converges, resulting in a fully trained detection model; In this context, the backbone network, encoder, density-guided target center heatmap generator, density-center feature enhancement module, decoder, and prediction head mentioned in steps S1 to S5 are all corresponding components in the trained detection model.

[0045] Specifically, the model training flowchart is as follows: Figure 3 As shown, the specific training steps are as follows: Input: Image Real Tag Set Hyperparameters and .

[0046] Step 1: Feature extraction. Extract the image... Input is fed into the backbone network, and the encoder extracts multi-scale features. .

[0047] Step 2: Density feature generation and target quantity estimation. Input the lowest layer feature map. To the quantity estimation branch, generate density features. And further input the counting and classification head to predict the number of targets. By comparing with the actual count labels Compare and calculate cross-entropy loss .

[0048] Step 3: Supervised heatmap generation. Lowest layer feature map. Preliminary target center features are obtained through heatmap branching. Then, the average value is calculated for each channel and normalized to obtain the corresponding heatmap. For each ground truth bounding box, a Gaussian kernel with density-adaptive variance is used to generate a target heatmap according to formulas (1) to (5). and calculate and Mean square error loss .

[0049] Step 4: Feature fusion and enhancement. This involves fusing and enhancing density features. Features of the target center The input is fed into the DEFE module and fused according to formulas (7) to (8) to obtain the optimized fused features. And will For multi-scale features Feature enhancement, resulting in an enhanced feature set. .

[0050] Step 5: Decoding and Prediction. This involves enhancing the features. Input multi-stage decoder, each intermediate stage The results generated are predicted by the head Obtain intermediate prediction results Used to calculate the corresponding auxiliary loss After summing, we get The decoder output in the final stage is obtained through prediction by the prediction head. ,in This is the number of decoder layers, then combined with... Joint calculation Finally, the overall training objective is calculated according to formula (9), and the model parameters are updated by minimizing the objective.

[0051] The overall loss function in this embodiment can be expressed as: (9) in, The weighting coefficient for the heatmap generation loss is set to 0.05 in the experiment. Corresponding to the heatmap generation loss introduced by the proposed DOCHG module, the module uses mean squared error (MSE) to supervise heatmap prediction, as defined in Equation (6). The original loss term in DQ-DETR is represented as follows: (10) in, It consists of three parts: (a) Hungarian loss (a) Counting loss (c) Auxiliary loss An additional Hungarian loss is introduced at each decoder stage to enhance training stability and accelerate convergence.

[0052] This embodiment achieves stable convergence and significantly improves the detection performance of densely distributed small targets in UAV images by jointly optimizing these loss terms in a unified training process.

[0053] The ablation experiment results of the model on the VisDrone dataset are shown in Table 1. The ablation experiment results on the VisDrone drone aerial photography dataset show that the introduction of these two modules can significantly improve the detection performance, verifying the effectiveness and complementarity of the two modules in improving the accuracy of dense small target recognition.

[0054] Table 1

[0055] Experimental results on two typical UAV aerial photography datasets, VisDrone and UAVDT (Tables 2 and 3) show that this embodiment achieves state-of-the-art performance across multiple evaluation metrics. Compared with existing general object detection methods and detection methods specifically designed for UAV scenarios, this embodiment demonstrates significant advantages in detecting small targets, particularly excelling in accurately locating densely packed small targets. This result fully validates the effectiveness of the density-center-guided perception enhancement strategy in improving detection accuracy and robustness in UAV aerial images, while also showcasing the method's strong adaptability to handling multi-scale targets and high-density scenes from a complex UAV perspective.

[0056] Table 2

[0057] Table 3

[0058] The experimental results of this embodiment on the VisDrone dataset are shown in Table 4, demonstrating significant advantages in both small and large-scale target detection. Specifically, it achieved the highest AP in the highly challenging "person" category, fully validating its superior ability to identify small targets in complex scenes with dense distribution, extremely small scale, and mutual occlusion. Simultaneously, it also achieved best performance in large target categories such as "car" and "truck," demonstrating good scale adaptability. Compared with existing methods, this embodiment outperforms or remains competitive in multiple categories, exhibiting strong robustness and generalization ability. Given that UAV aerial photography scenes often contain both small-scale dense targets and large-scale sparse targets, the balanced performance of this method across different categories makes it particularly suitable for target detection tasks in complex UAV scenes.

[0059] Table 4

[0060] The beneficial effects of this embodiment: This embodiment addresses the challenge of detecting dense small targets in UAV aerial images by proposing a density-center-guided perception enhancement strategy. By combining global density features with target center features, the model can accurately distinguish adjacent small targets in dense regions, significantly reducing the false negative rate of overlapping targets. Compared with existing methods that rely solely on global features or region cropping, this embodiment fully utilizes the complementarity between global information and target local information, particularly improving the ability to identify small, closely packed targets, thus making it more suitable for target detection tasks in complex UAV scenarios.

[0061] This embodiment proposes two core modules: a density-guided target center heatmap generator (DOCHG) and a density-center feature enhancement module (DCFE). DOCHG generates a density-aware target center heatmap, guiding the model to focus on the center position of each target, effectively mitigating the boundary blurring and overlap problems that easily occur with dense small targets in UAV aerial images. DCFE fuses density features with the center heatmap, further enhancing the ability to represent target features, enabling the model to accurately distinguish adjacent small targets in complex UAV scenes.

[0062] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A density center-guided perception enhancement method for detecting dense small targets in UAV imagery, characterized in that, Includes the following steps: S1: Acquire drone aerial images, input the drone aerial images into a preset backbone network and encoder, and extract multi-scale feature maps; S2: Select a specified layer feature map from the multi-scale feature map, input the specified layer feature map into the density-guided target center heatmap generator, and generate density features through the quantity estimation branch. The density-guided target center heatmap generator combines the local target density information contained in the density features to generate a density-adaptive target center heatmap for supervising the generation of target center features. S3: Input the density feature and the target center feature into the density-center feature enhancement module. The density-center feature enhancement module first performs dilated convolution processing on the density feature to expand the receptive field, and performs spatial attention weight generation processing on the target center feature to highlight the target center region. Then, it calculates the similarity between the processed density feature and the processed target center feature to obtain a fusion weight map. Finally, based on the fusion weight map, it performs weighted fusion of the two types of processed features and performs channel weighting to obtain the enhanced feature. S4: Fuse the enhanced features with the multi-scale feature map to obtain a fused feature set; S5: Input the fused feature set into a preset decoder and prediction head, and output the detection result of the target in the UAV image after processing by the decoder and prediction head.

2. The method according to claim 1, characterized in that, In step S1, the drone aerial images are input into a preset backbone network and encoder to extract multi-scale feature maps, specifically including: The drone aerial images are input into the backbone network, and initial features at different resolutions are obtained after feature extraction layer by layer by the backbone network. Each of the initial features is input into the encoder. After feature encoding and scale alignment by the encoder, the multi-scale feature map containing feature information at different scales is output.

3. The method according to claim 1, characterized in that, Step S2 involves selecting a specified layer feature map from the multi-scale feature map, specifically including: The spatial detail retention of each layer of the multi-scale feature map is analyzed, and the layer feature map with the most complete spatial detail retention and the ability to reflect the local information of the small target is selected as the specified layer feature map.

4. The method according to claim 1, characterized in that, The density-guided target center heatmap generator in step S2 combines the local target density information contained in the density features to generate a density-adaptive target center heatmap, specifically including: The density-guided target center heatmap generator calculates the local target density within each real target bounding box based on the density features. The variance parameter of the Gaussian kernel is adjusted according to the local target density to generate a Gaussian response at the center of the real target bounding box. A coefficient related to the target area is introduced to enhance the response of small targets in the heat map, and finally the density-adaptive target center heat map is generated iteratively.

5. The method according to claim 1, characterized in that, Step S3 calculates the similarity between the processed density features and the processed target center features to obtain the fused weight map, specifically including: Convolution operations are performed on the processed density features and the processed target center features respectively to keep the number of channels of the two types of features consistent. For two types of features with the same number of channels, multiply them element by element and sum the results along the channel dimension; The summation result is input into the activation function for normalization to obtain the fusion weight map that reflects the similarity between the two types of features.

6. The method according to claim 1, characterized in that, Step S3 involves weighted fusion of the two processed features based on the fusion weight map, specifically including: The fusion weight map is multiplied element-wise with the processed density features to obtain the first fusion part dominated by the density features; Calculate the difference between 1 and the fusion weight map, and multiply the difference element-wise with the processed target center feature to obtain the second fusion part dominated by the target center feature; The first fusion portion and the second fusion portion are added element by element to obtain a fusion feature that takes into account both regional density information and target center information.

7. The method according to claim 1, characterized in that, Step S3 involves channel weighting of the weighted fusion result, specifically including: The weighted fusion results are subjected to feature statistics at the channel dimension, and the contribution of each channel feature to target detection is calculated. Assign weight coefficients to the corresponding channels based on their respective contributions; Each channel feature is multiplied element-wise with its corresponding weight coefficient to obtain the enhanced features that strengthen key features and suppress redundant features.

8. The method according to claim 1, characterized in that, Step S4 involves fusing the enhanced features with the multi-scale feature map to obtain a fused feature set, specifically including: The enhanced features are scaled so that the resolution of the enhanced features is consistent with the resolution of the corresponding layer feature map in the multi-scale feature map; The enhanced features with consistent resolution and the corresponding layer feature maps are processed by spatial attention and channel attention mechanisms to obtain fused features of each layer that integrate density-center enhancement information. The fusion features of all layers are summarized to form the fusion feature set.

9. The method according to claim 1, characterized in that, Step S5, after processing by the decoder and prediction head, outputs the target detection results in the UAV imagery, specifically including: The fused feature set is input into the decoder, which performs multi-stage decoding on the fused feature set in conjunction with a preset dynamic query, and outputs intermediate predicted features at each stage to optimize the decoding process. The final stage decoding features are input into the prediction head, which performs target category prediction and bounding box regression on the final stage decoding features, and outputs the detection result containing target category and location information.

10. The method according to claim 1, characterized in that, It also includes the model training step: Obtain a training dataset, which contains multiple drone aerial sample images labeled with target categories and bounding boxes; Based on the training dataset, and with a preset total loss function as the optimization objective, the overall model, consisting of the backbone network, encoder, density-guided target center heatmap generator, density-center feature enhancement module, decoder, and prediction head, is iteratively trained. Continue until the model converges, resulting in a fully trained detection model; In this context, the backbone network, encoder, density-guided target center heatmap generator, density-center feature enhancement module, decoder, and prediction head mentioned in steps S1 to S5 are all corresponding components in the trained detection model.

Images