Unmanned aerial vehicle small target detection method based on edge information flow guidance

Abstract

The invention discloses an unmanned aerial vehicle small target detection method based on edge information flow guidance, and relates to the technical field of target detection, and the method comprises the steps: obtaining an input image, and constructing an unmanned aerial vehicle small target detection training set; constructing a small target detection model according to the backbone network, the ino-PAN structure and the decoding network which are connected in sequence; training a small target detection model through the unmanned aerial vehicle small target detection training set; and obtaining an unmanned aerial vehicle real-time image and inputting the image into the trained small target detection model to obtain an unmanned aerial vehicle small target detection result. According to the method, Gaussian edge features are fused with sampling information of different strategies, so that information of more small-size targets is obtained. By supplementing the edge feature information to the fusion network, the loss of the edge information in the fusion process is reduced. Through weight control feature information fusion, low-efficiency feature fusion into a network is reduced, and effective features are better utilized.

Description

Technical Field

[0001] This invention relates to the field of target detection technology, and more specifically to a method for detecting small targets on unmanned aerial vehicles (UAVs) based on edge information flow guidance. Background Technology

[0002] In recent years, the rapid development of computer vision technology has greatly promoted the progress of visual perception systems. Against this backdrop, unmanned aerial vehicles (UAVs), with their high-altitude maneuverability, have seen a continuous improvement in their ability to acquire visual information in complex environments. Images captured by UAVs typically cover vast areas, many of which contain small-sized key targets, making small target detection a crucial supporting technology for improving image processing efficiency. This technology has demonstrated significant value in tasks such as forestry surveys, urban monitoring, and disaster relief. However, the inherent high angle of view and wide field of view characteristics of UAV photography lead to a significant reduction in the pixel ratio of targets, especially resulting in extremely sparse feature representation for small targets. Simultaneously, complex background interference (such as vegetation and buildings) and drastic scale changes caused by perspective effects pose serious challenges to the accuracy and robustness of current mainstream detection algorithms, severely restricting the intelligent application level of UAV systems in real-world scenarios. Therefore, developing efficient small target detection algorithms for UAV scenarios has become a key issue for improving perception performance and ensuring mission reliability.

[0003] Driven by deep learning, significant progress has been made in the field of general object detection. Classic methods, such as the YOLO series and Faster R-CNN, as well as novel detectors based on the Transformer architecture, such as DETR, SwinTransformer, RT-DETR, and DINO, have all demonstrated strong performance on multiple general benchmark datasets. However, these algorithms still face significant bottlenecks when transferred to the scenario of small object detection on UAVs. Specifically, on the one hand, during deep feature extraction, the semantic information of small objects is easily attenuated or even lost during convolution operations, making it difficult for traditional backbone networks to effectively capture fine-grained features such as edge textures. On the other hand, although existing multi-scale fusion mechanisms such as FPN can integrate feature information from different levels, they often lack targeted enhancement strategies for small object features, making them easily suppressed by large objects or background features. These factors together lead to poor performance of algorithms in small object detection on UAVs. In addition, attitude changes during the flight of the UAV platform, such as pitch and rotation, further exacerbate the deformation and occlusion effects of the target, further reducing detection accuracy.

[0004] To address the core challenges of small target detection on UAVs, such as feature decay, insufficient multi-scale fusion, and interference from attitude changes, researchers have explored various improvement paths and achieved some success. Regarding enhancing feature representation capabilities, for example, Liu et al. effectively enhanced the representational ability of shallow features for small targets by introducing additional high-resolution feature layers or constructing more refined pyramid structures. However, these methods typically involve a significant increase in computational complexity, making it difficult to meet the real-time processing requirements of UAV platforms. To further enhance feature discriminative power, Li et al. utilized attention mechanisms to focus on key regions to enhance the feature response of small targets, or established inter-target relationships through contextual modeling techniques similar to Relation Networks, improving target recognition capabilities in complex backgrounds. While these methods can improve detection performance to some extent, their depth and robustness in modeling the distribution patterns and intrinsic relationships of small targets in complex scenes remain insufficient, limiting performance improvement. In recent years, more exploratory generative detection methods, such as Diffusion Det, have also been introduced into this field. These methods, by simulating the data generation process, theoretically possess stronger capabilities in modeling complex distributions and handling deformation. However, its reliance on multiple iterations of optimization makes it prone to convergence difficulties or efficiency bottlenecks when dealing with small targets, thus limiting its practical application value.

[0005] Therefore, how to propose a small target detection method for UAVs based on edge information flow guidance to overcome the shortcomings of the existing technology is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] In view of this, this invention provides a method for small target detection in UAVs based on edge information flow guidance, addressing the challenges of extremely low pixel count and sparse features of small targets in UAV high-altitude photography, as well as the problem that existing algorithms easily lose or suppress small target information during feature extraction and fusion. Specifically, a Gaussian Edge Enhancement (GEE) module is seamlessly integrated into the backbone network, which effectively promotes the extraction of edge features crucial for accurate target identification. Following the neck of the small-scale branch, a Small Target Feature Enhancement (SFB) module is designed. The SFB module dynamically enhances the feature representation of small targets, effectively alleviating the inherent difficulties in their detection. In the information interaction stage, an information flow path aggregation network structure is proposed. This structure bridges the gap between deep semantic features and shallow detail features through an information flow mechanism, unifying multi-scale feature representation, thereby significantly improving the scale adaptability and localization accuracy of target detection. To achieve the above objectives, this invention adopts the following technical solution: A method for detecting small targets on a drone based on edge information flow guidance, comprising: Acquire the input image and construct a training set for UAV small target detection; A small target detection model is constructed based on the sequentially connected backbone network, info-PAN structure, and decoding network. The small target detection model was trained using a small target detection training set for UAVs. The real-time images of the UAV are input into the trained small target detection model to obtain the small target detection results of the UAV.

[0007] Optionally, the backbone network consists of multiple interconnected convolutional blocks and Gaussian edge enhancement modules.

[0008] Optionally, the backbone network specifically includes: a dual convolutional block, a Gaussian edge enhancement module and a group of convolutional blocks, a Gaussian edge enhancement module, a convolutional block, a Gaussian edge enhancement module, a convolutional block, and a Gaussian edge enhancement module connected in sequence.

[0009] Optionally, the Gaussian edge enhancement module includes: an input module, a convolutional block, and a channel segmentation module connected in sequence, which are divided into two branches by the channel segmentation module; the features of the first branch undergo two different operations, the first operation being a convolutional block, and the second operation being an adaptive average pooling module, a GE module, and an interpolation module; then the outputs of these two operations are concatenated by the channel splicing module and then coupled through the convolutional block for output; the coupled output result is concatenated with the output result of the second branch by the channel splicing module and then passed through the convolutional block and the output module for final output.

[0010] Optionally, the GE module includes: a convolutional block input, which is subjected to a subtraction operation after passing through a convolutional block and a Gaussian blur module, and the subtraction result is output through a convolutional block.

[0011] Optional formulas for calculating Gaussian blur include: ; ; Where (i, j) are the center coordinates, c is the kernel size, σ is the standard deviation, Z is the normalization constant, x is the input feature map, and B(x) is the result after Gaussian blurring. Represents the Gaussian kernel.

[0012] Optionally, the info-PAN structure includes: Construct a pyramid structure with layers P2, P3, P4, and P5; Edge information is extracted from the features of layer P2 by the information flow module to form an edge information flow, which is then added to the pyramid structure. Layers P5, P4, and P3 sequentially combine edge information for progressive upsampling and downsampling, ultimately outputting multiple results to the decoding network.

[0013] Optionally, the info-PAN structure specifically includes: Edge information is extracted from the features of layer P2 through the information flow module to form edge information of layers P4, P3 and P5; In the upsampling branch, the input of P5 layer is fed into P4 layer via a convolutional block; in P4 layer, the output of P5 layer, the input of P4 layer, and the edge information of P4 layer are coupled through the SFB module and output to P3 layer; in P3 layer, the output of P4 layer, the input of P3 layer, and the edge information of P3 layer are coupled through the SFB module and output to the downsampling branch. In the downsampling branch, the output of the upsampling branch of layer P3 is received in layer P3 and processed by a convolutional block before being input to the decoding network and layer P4. In layer P4, the output of the downsampling branch of layer P3, the output of the upsampling branch of layer P4, and the edge information of layer P4 are coupled through the SFB module and output to layer P5 and the decoding network. In layer P5, the output of the downsampling branch of layer P4, the output of the upsampling branch of layer P5, and the edge information of layer P5 are coupled through the SFB module and output to the decoding network.

[0014] Optionally, the information flow module includes: a GE module and three convolutional blocks connected in sequence. The GE module receives the input of the P2 layer, and the three convolutional blocks output the edge information of the P4, P3 and P5 layers respectively.

[0015] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a method for detecting small targets of UAVs based on edge information flow guidance, which has the following beneficial effects: This invention proposes a method for small target detection in unmanned aerial vehicles (UAVs) based on edge information flow guidance, comprising: acquiring an input image and constructing a UAV small target detection training set; constructing a small target detection model based on a backbone network, an info-PAN structure, and a decoding network connected in sequence; training the small target detection model using the UAV small target detection training set; and acquiring real-time images of the UAV and inputting them into the trained small target detection model to obtain the UAV small target detection result. This invention proposes a Gaussian edge enhancement sampling module, which obtains more information about small targets by fusing Gaussian edge features with sampling information from different strategies. An Info-PAN structure is designed to supplement the fusion network with edge feature information, reducing edge information loss during the fusion process. A soft-connection fusion module is proposed, which controls feature information fusion through weights, reducing the inefficient integration of features into the network and allowing effective features to be better utilized. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0017] Figure 1 This invention provides a structural principle diagram of a UAV small target detection method based on edge information flow guidance.

[0018] Figure 2 A comparative diagram of the Info-PAN structure provided by this invention.

[0019] Figure 3 The information flow module structure diagram provided by this invention.

[0020] Figure 4 The diagram shows the structure of the SFB module provided by this invention.

[0021] Figure 5 The ablation analysis results of the GEE module provided by this invention are shown in the figure.

[0022] Figure 6 The ablation analysis results of the Info-PAN structure provided by this invention are shown in the figure.

[0023] Figure 7 This is a visualization of the model detection results on the VisDrone dataset provided by the present invention.

[0024] Figure 8 This is a visualization of the model detection results on the TT100k dataset provided by the present invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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 discloses a method for detecting small targets on unmanned aerial vehicles (UAVs) based on edge information flow guidance, comprising: Acquire the input image and construct a training set for UAV small target detection; A small target detection model is constructed based on the sequentially connected backbone network, info-PAN structure, and decoding network. The small target detection model was trained using a small target detection training set for UAVs. The real-time images of the UAV are input into the trained small target detection model to obtain the small target detection results of the UAV.

[0027] Furthermore, the backbone network consists of multiple interconnected convolutional blocks and Gaussian edge enhancement modules.

[0028] Furthermore, the backbone network specifically includes: a dual convolutional block, a Gaussian edge enhancement module and a group of convolutional blocks connected in sequence, a Gaussian edge enhancement module, a convolutional block, a Gaussian edge enhancement module, a convolutional block, and a Gaussian edge enhancement module.

[0029] Furthermore, the Gaussian edge enhancement module includes: an input module, a convolutional block, and a channel segmentation module connected in sequence. The channel segmentation module divides the feature into two branches. The first branch features undergo two different operations: the first operation is a convolutional block, and the second operation is through an adaptive average pooling module, a GE module, and an interpolation module. The outputs of these two operations are then concatenated by a channel splicing module and coupled through a convolutional block for output. The coupled output is then concatenated with the output of the second branch through a channel splicing module, and finally passed through a convolutional block and an output module for output. Further, the GE module includes: a convolutional block input, which undergoes a subtraction operation after passing through a convolutional block and a Gaussian blur module, and the subtraction result is output through a convolutional block.

[0030] Furthermore, the formula for calculating Gaussian blur includes: ; ; Where (i, j) are the center coordinates, c is the kernel size, σ is the standard deviation, Z is the normalization constant, x is the input feature map, and B(x) is the result after Gaussian blurring. Represents the Gaussian kernel.

[0031] Furthermore, the info-PAN structure includes: Construct a pyramid structure with layers P2, P3, P4, and P5; Edge information is extracted from the features of layer P2 by the information flow module to form an edge information flow, which is then added to the pyramid structure. Layers P5, P4, and P3 sequentially combine edge information for progressive upsampling and downsampling, ultimately outputting multiple results to the decoding network.

[0032] Furthermore, the info-PAN structure specifically includes: Edge information is extracted from the features of layer P2 through the information flow module to form edge information of layers P4, P3 and P5; In the upsampling branch, the input of P5 layer is fed into P4 layer via a convolutional block; in P4 layer, the output of P5 layer, the input of P4 layer, and the edge information of P4 layer are coupled through the SFB module and output to P3 layer; in P3 layer, the output of P4 layer, the input of P3 layer, and the edge information of P3 layer are coupled through the SFB module and output to the downsampling branch. In the downsampling branch, the output of the upsampling branch of layer P3 is received in layer P3 and processed by a convolutional block before being input to the decoding network and layer P4. In layer P4, the output of the downsampling branch of layer P3, the output of the upsampling branch of layer P4, and the edge information of layer P4 are coupled through the SFB module and output to layer P5 and the decoding network. In layer P5, the output of the downsampling branch of layer P4, the output of the upsampling branch of layer P5, and the edge information of layer P5 are coupled through the SFB module and output to the decoding network.

[0033] Furthermore, the information flow module includes: a GE module and three convolutional blocks connected in sequence. The GE module receives the input of the P2 layer, and the three convolutional blocks output the edge information of the P4, P3 and P5 layers respectively.

[0034] In specific implementations, to address the challenges of small object recognition, insufficient feature information interaction, and inadequate fusion of deep and shallow features in multi-scale target detection, this invention proposes a detection network called GFS-DETR (Gaussian-edge Fusion Selection DEtection TRansformer), such as... Figure 1 As shown, specifically, the GEE module is seamlessly integrated into the backbone network, facilitating the extraction of edge features crucial for accurate identification. Following the neck portion of small-scale branches, an SFB module is designed to dynamically enhance the feature representation of small objects and address the inherent challenges associated with their detection. Furthermore, the Info-PAN structure is strategically placed in large-scale branches to enhance feature information interaction. The PAN structure bridges the gap between deep semantics and shallow details through Information Flow, unifying multi-scale feature representation and significantly enhancing the scale adaptability and localization accuracy of object detection. Through the synergistic effect of these tailored modules, significant enhancements in detection accuracy are demonstrated, highlighting its effectiveness and versatility in addressing the challenges of multi-scale object detection, not only for small objects but also for large objects.

[0035] In specific implementations, in the context of object detection, the limited number of pixels occupied by small objects poses a unique challenge, leading to a greater likelihood of feature loss during the feature extraction stage of the backbone network. Current detection methods primarily rely on shallow maps to detect these small objects, which also include redundant background information. This redundancy not only increases computational demands during the detection stage but also slows down the model's inference speed. Considering this limitation, a Gaussian Edge Enhancement Module (GEE module) is proposed to inject edge information into the extracted features, enhancing the feature representation for small object detection. The GEE module utilizes edge information to guide the model's learning of small object features, thereby enhancing the discriminability of small object features. Gaussian edge information contributes to the saliency of small object features, ultimately improving the accuracy of detecting these small objects during feature fusion. The structure of the GEE module is as follows: Figure 1 As shown on the right. Image edge information typically appears in areas of rapid pixel change, i.e., locations with large pixel gradient variations. Gaussian blur can smooth the input feature map by applying a low-pass filter to separate high-frequency features. Then, edge features are obtained by subtracting the low-pass filtered feature map from the original feature map.

[0036] The formula for calculating Gaussian blur is as follows: ; ; Where (i, j) are the center coordinates, c is the kernel size, σ is the standard deviation, Z is the normalization constant, x is the input feature map, and B(x) is the result after Gaussian blurring. Represents the Gaussian kernel.

[0037] Based on the above theory, in the GEE module, a 5×5 Gaussian blur window is used to smooth the input features and extract low-frequency information from the feature map. Then, the collected low-frequency information is subtracted from the original feature map to obtain the edge information. As shown in the following formula: ; Where x is the input feature map, It's a convolution, and B(x) is the result after Gaussian blurring. It is edge information.

[0038] Specifically, such as Figure 1As shown, to further improve the target detection performance of blurred feature images under harsh conditions, features are extracted through pooling layers using Adaptive Avg Pool. This allows the network to capture more comprehensive information. After convolving feature maps at different scales, the edge contour features and details of targets in the image are enhanced using the GE module, and then the edge feature maps are concatenated with the residuals. This compensates for the problems of indistinct features of small targets and insufficient edge contour extraction, effectively improving the network's detection performance of weakly textured and weakly edged targets caused by UAVs at different angles and under different lighting conditions.

[0039] In a specific implementation, to address the challenges posed by varying angles and lighting conditions in UAV aerial imagery, which make identification difficult, an Info-PAN structure is proposed. The structure of this model is as follows: Figure 2 As shown, this approach aims to address the challenges of difficult feature extraction and the loss of small target features. First, a PAN structure is used to interact the feature information from layers P3, P4, and P5. Second, outside the fusion structure, edge information is extracted from the features of layer P2 to form an edge information stream. Finally, this edge information is added to the PAN fusion structure. This allows additional edge information to compensate for the small target features lost during the fusion process. The Info-PAN structure is shown below. Figure 2 As shown in (c), compared to FPN and BiFPN, Info-PAN solves the problem of FPN's simple deep-to-shallow fusion, where additional information makes the detected features lack original information.

[0040] Specifically, the GE module within the GEE module is used to collect edge information from the P2 feature layer, and convolutional downsampling is used to obtain the corresponding layer information, forming information streams for each layer. The information stream module is as follows: Figure 3 As shown, edge information is supplemented into the fusion network to solve the problem of edge information loss during the fusion process. At the same time, it also solves the problem of the BiFPN structure causing a surge in computation and thus slowing down the detection speed.

[0041] In specific implementations, to filter low-quality features during information interaction, a Soft-combination Fusion Block (SFB) is proposed, such as... Figure 4 As shown, since feature maps from different sources differ in the number of channels, stride, and receptive field, simple concatenation or weighted averaging will reduce feature quality. The SFB module achieves more accurate fusion by implementing feature alignment and adaptive weighting, thereby improving the performance of the object detection network.

[0042] Specifically, given a set of input feature maps Each feature map Since they may have different numbers of channels and spatial dimensions, the SFB module first performs channel alignment. Let the unified number of channels be C1, the number of channels in the first feature map. Then the alignment process can be expressed as: ; in, It is the number of channels in the first feature map. This refers to the number of channels. Features Use convolution to convert to the number of channels. .

[0043] Subsequently, the SFB module generates adaptive weights through an attention mechanism. First, it performs global average pooling on each aligned feature map, and then generates weight scores through a two-layer fully connected network. Finally, it performs softmax normalization on all weight scores, as shown in the following formula: ; ; ; GAP stands for Global Average Pooling. This represents a fully connected layer used to generate importance scores for features; Represents the unnormalized importance score of the i-th feature; It is the feature weights after normalization by the softmax function; the final feature representation after F-weighted fusion.

[0044] The SFB module fuses features from different algebras using the strategy described above, maintaining a channel count of C1, and ultimately outputs the feature F after multi-layer fusion. This design ensures that the SFB module can dynamically adjust the contribution weights of each feature map, effectively filtering out low-quality features, while preserving original information through residual connections. This significantly improves feature fusion quality and detection performance in multi-scale object detection tasks.

[0045] In a specific implementation, the fused features undergo another feature enhancement operation. The strategy is to divide the features into a 1:3 ratio, and then process each segmented feature through a convolutional process. Features with a smaller proportion undergo an additional Rep convolutional process, ultimately resulting in the fused features. The calculation formula is as follows.

[0046] ; ; in, It is a channel segmentation function that segments feature F into channels. The channel ratio is 1:3; RepConv represents Rep convolution, and Conv represents convolution. Channel splicing representing features.

[0047] Augmentation operations enable the model to enhance its representational capacity and multi-scale feature extraction capabilities during the training phase, thereby improving its discriminative performance. During the inference phase, structural reparameterization techniques are used to fuse multiple branches into a single, efficient convolution, which significantly improves forward inference speed while maintaining high accuracy, achieving an effective balance between model expressiveness and computational efficiency.

[0048] In a specific implementation, a method for detecting small targets on a drone based on edge information flow guidance includes the following experimental steps: (1) The network experimental environment was based on Ubuntu 20.04, Python 3.8.19, and PyTorch 2.4.1. The computer used in the experiment was equipped with an Intel i9-12900K processor, 125GB of memory, an RTX 3090 graphics card with 24GB of video memory. The training batch size was set to 4, the number of training epochs was set to 300, and the learning rate was set to 5×10. 3. Select an adaptive image size of 640×640 for the experiment. Other experimental settings are consistent with RT-DETR.

[0049] (2) The VisDrone2019-DET dataset is a drone aerial image target detection dataset constructed by the AISKYEYE team of Tianjin University. This dataset is designed specifically for understanding visual data captured by drones and covers complex environments and challenges such as different scenes and different lighting conditions. The dataset contains 7019 images, of which the training set VisDrone 2019-train contains 6471 images and the validation set VisDrone 2019-val contains 548 images. The dataset has 10 predefined target categories: pedestrians, people, bicycles, cars, vans, trucks, tricycles, covered tricycles, buses, and motorcycles.

[0050] TT100K (Tsinghua-Tencent 100K) is a large-scale traffic sign detection and recognition dataset jointly released by Tsinghua University and Tencent. It contains over 100,000 images of Chinese road scenes, of which more than 30,000 are annotated, covering 221 categories of traffic signs, with a total of over 300,000 annotated targets. The data mainly consists of high-resolution (2048×2048) images, and the annotation information provides precise bounding boxes, covering multiple types of signs such as prohibitions, warnings, and instructions.

[0051] (3) In this embodiment, by maintaining consistent experimental conditions, the performance changes in image detection before and after model optimization are compared to measure the efficiency of the algorithm. The evaluation is based on the following criteria: precision, recall, average precision (mAP), and floating-point operations per second (GFLOPs).

[0052] (4) Ablation experiments: To further confirm the efficacy of the method proposed in this embodiment, a series of ablation studies on the Visdrone and TT100k datasets were designed and executed. These rigorous experiments were specifically tailored to analyze and evaluate the individual contributions of each component in the framework of this embodiment.

[0053] 1) Overall component ablation analysis: To establish a basis for comparison, RT-DETR was selected as the baseline model, and different modules were gradually introduced to systematically enhance it. Specific experimental configurations included: (1) using only the baseline model; (2) the baseline combined with the GEE module; (3) the baseline combined with the GEE module and the Info-PAN structure; and (4) a fully integrated model, i.e., adding the GEE, Info-PAN, and SFB modules to the baseline simultaneously. Ablation experiments on the Visdrone and TT100k datasets allowed for clear isolation and quantification of the contribution of each module to the overall performance. The quantitative results of these experiments are listed in Tables 1 and 2, respectively. On the Visdrone dataset, compared to the baseline, our model improved by 3.5% on mAP50 and 2.4% on the mAP50-95 metric. In the evaluation on the TT100k dataset, the model improved by 2.5% on mAP50 compared to the baseline. Notably, the target achieved a significant improvement of 1.8% on the mAP50-95 metric. In addition, Recall and Precision improved by 2.1% and 4.0% respectively on targets of different sizes.

[0054] Table 1. Ablation experiment results of the algorithm module on VisDrone

[0055] Table 2 Ablation experiment results of the algorithm module on TT100K

[0056] 2) Ablation analysis of the GEE module: To rigorously evaluate the ability of the GEE module to enhance the detection performance of large-scale objects by extending the receiver field, a comprehensive set of ablation experiments were conducted on the Vistrone and TT100k datasets. The empirical results of these experiments, listed in Table 3, provide compelling evidence that the integration of the SFB module leads to a significant improvement in the detection accuracy of objects at different scales, thus demonstrating its effectiveness and practicality. Figure 5 The experimental results clearly demonstrate that the incorporation of the GEE module significantly enhances the overall detection accuracy of our model. This improvement stems from the GEE module's ability to model using global information, effectively expanding the model's receptive field. Therefore, this enhancement contributes to a more comprehensive understanding of the input data, ultimately leading to improved model performance and detection capabilities.

[0057] Table 3. Comparison of ablation results with and without the GEE module on the VisDrone and TT100k datasets.

[0058] 3) Ablation analysis of the Info-PAN structure: To rigorously evaluate the role of the Info-PAN structure in enhancing small target features, comprehensive ablation experiments were conducted on the Visdrone and TT100k datasets, and the results are shown in Table 4. Quantitative analysis clearly demonstrates that the introduction of the Info-PAN structure significantly improves the detection performance of small targets, thus highlighting its crucial role in the overall framework. By achieving dynamic fusion between feature pipelines at different scales, the Info-PAN structure effectively preserves rich features related to the edges of small targets, thereby greatly enhancing the detection capability for such targets.

[0059] To more intuitively demonstrate the effectiveness of the Info-PAN structure in feature extraction, a comparative visualization analysis was conducted on feature maps with and without this structure. For example... Figure 6 As shown, the results indicate that when the model integrates the Info-PAN structure, it can more fully retain the detailed information of small targets, thereby achieving a more discriminative mapping within limited target features and significantly enhancing the detection capability of small targets.

[0060] Table 4. Comparison of ablation results with and without Info-PAN structures on the VisDrone and TT100k datasets.

[0061] 4) Ablation analysis of the SFB module: To evaluate the ability of the SFB module to improve the detection performance of large-scale targets, systematic ablation experiments were also conducted on the Visdrone and TT100k datasets, and the results are shown in Table 5. The experimental results provide strong evidence that the model's detection accuracy on targets of different scales is significantly improved after introducing the SFB module, thus verifying the effectiveness and practical value of the module. Specifically, the SFB module significantly expands the model's receptive field through global information modeling, enabling it to understand the input data more comprehensively. This improvement not only enhances the overall detection accuracy but also strengthens the model's ability to recognize large-scale targets.

[0062] Table 5. Comparison of ablation results with and without the SFB module on the VisDrone and TT100k datasets.

[0063] 5) To rigorously verify the effectiveness of the proposed model, several state-of-the-art UAV target detection methods were selected and compared on the Visdrone2019 dataset. The comparison methods included representative models such as Faster R-CNN, RetinaNet, YOLO series, and RT-DETR series. The experimental results are shown in Table 6, systematically demonstrating the performance of current mainstream detection methods on this benchmark dataset. Through comparison with these methods, the advantages of the proposed GFS-DETR model in target detection are clearly evident, further highlighting its potential value in UAV image detection tasks.

[0064] Table 6 Comparative Test Table

[0065] 6) Quantitative comparison: To rigorously verify the effectiveness of the proposed model, six state-of-the-art UAV target detection methods were selected on the Visdrone2019 dataset for comparative analysis. The comparison methods included representative models such as Faster R-CNN, RetinaNet, YOLO series, and RT-DETR series. The experimental results are shown in Table 6, systematically demonstrating the performance of current mainstream detection methods on this benchmark dataset. Through comparison with these methods, the advantages of the proposed GFS-DETR model in target detection are clearly evident, further highlighting its potential value in UAV image detection tasks.

[0066] 7) Qualitative comparison: To verify the performance of the GFS-DETR model, comparative experiments were conducted on the Visdrone and TT100k datasets. The detection results are visualized as follows: Figure 7 and Figure 8 As shown in the images, these visualizations visually demonstrate the model's capabilities in detecting small and multi-scale objects, significantly outperforming other methods. It is worth noting that, as... Figure 7 As shown, the blue dashed boxes mark targets that other models failed to detect, while GFS-DETR successfully identified these overlooked objects. In a direct comparison with RT-DETR, the model in this embodiment demonstrates higher detection accuracy on the Visdrone dataset, especially in scenarios with small targets and densely distributed multi-scale targets, showcasing the model's robustness and effectiveness.

[0067] Compared to the Visdrone dataset, the TT100k dataset presents a greater challenge due to its significant variations in object scale and higher scene complexity. To comprehensively evaluate the robustness and generalization ability of GFS-DETR, additional qualitative comparative experiments were conducted on the TT100k dataset. Figure 8 As shown in the visualization, the results clearly demonstrate that the model can still efficiently detect targets in complex scenes. This performance improvement is largely attributed to the introduction of the Gaussian Edge Enhancement (GEE) module, which effectively enhances feature extraction, enabling the model to maintain excellent detection capabilities in complex environments. Overall, through performance analysis on different datasets, the model in this embodiment can provide stable and reliable detection results in various complex scenarios, fully demonstrating its applicability in real-world applications.

[0068] In summary, existing target detection models often neglect edge information during design, posing a significant challenge to small target detection in UAV scenarios. To address this, this invention proposes a novel edge-guided Transformer detector that effectively improves the detection accuracy of small targets by introducing edge information. First, to overcome the inefficiency of model information acquisition, a Gaussian edge enhancement module is designed to increase the extraction of target edge features, ensuring effective utilization of edge information. By strengthening edge features related to small targets and reallocating fusion weights during the feature fusion stage, the model's sensitivity and representation ability for small targets are effectively improved. Furthermore, an Info-PAN structure is proposed, which bridges the gap between deep semantics and shallow details through an information flow mechanism, effectively solving the problem of lost edge information for small targets during fusion. Finally, to address the shortcomings of the fusion module, a soft-combination fusion block is designed to dynamically fuse features of different sizes and enhance the feature representation of small targets, thereby alleviating the problem of insufficient information for small targets. Experimental results on the VisDrone dataset demonstrate that the proposed model outperforms current mainstream models in both mean accuracy (mAP50) and mAP50-95. Further ablation experiments validate the effectiveness of the introduced modules in improving target detection performance in UAV scenarios.

[0069] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0070] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for detecting small targets on unmanned aerial vehicles (UAVs) based on edge information flow guidance, characterized in that, include: Acquire the input image and construct a training set for UAV small target detection; A small target detection model is constructed based on the sequentially connected backbone network, info-PAN structure, and decoding network. The small target detection model was trained using a small target detection training set for UAVs. The real-time images of the UAV are input into the trained small target detection model to obtain the small target detection results of the UAV.

2. The method for detecting small targets on a UAV based on edge information flow guidance according to claim 1, characterized in that, The backbone network consists of multiple interconnected convolutional blocks and Gaussian edge enhancement modules.

3. The method for detecting small targets on a UAV based on edge information flow guidance according to claim 2, characterized in that, The backbone network specifically includes: a dual convolutional block, a Gaussian edge enhancement module and a group of convolutional blocks, a Gaussian edge enhancement module, a convolutional block, a Gaussian edge enhancement module, and a convolutional block, all connected in sequence.

4. The method for detecting small targets on a UAV based on edge information flow guidance according to claim 3, characterized in that, The Gaussian edge enhancement module includes an input module, a convolutional block, and a channel segmentation module connected in sequence. The channel segmentation module divides the features into two branches. The features of the first branch undergo two different operations: the first operation is a convolutional block, and the second operation is through an adaptive average pooling module, a GE module, and an interpolation module. The outputs of these two operations are then concatenated by a channel splicing module and coupled through a convolutional block before being output. The coupled output and the output of the second branch are then concatenated by a channel splicing module, passed through a convolutional block and an output module, and finally output.

5. The method for detecting small targets on a UAV based on edge information flow guidance according to claim 4, characterized in that, The GE module includes: a convolutional block input, which is subjected to a subtraction operation after passing through a convolutional block and a Gaussian blur module, and the subtraction result is output through a convolutional block.

6. The method for detecting small targets on a UAV based on edge information flow guidance according to claim 5, characterized in that, The formulas for calculating Gaussian blur include: ； ； Where (i, j) are the center coordinates, c is the kernel size, σ is the standard deviation, Z is the normalization constant, x is the input feature map, and B(x) is the result after Gaussian blurring. Represents the Gaussian kernel.

7. The method for detecting small targets on a UAV based on edge information flow guidance according to claim 1, characterized in that, The info-PAN structure includes: Construct a pyramid structure with layers P2, P3, P4, and P5; Edge information is extracted from the features of layer P2 by the information flow module to form an edge information flow, which is then added to the pyramid structure. Layers P5, P4, and P3 sequentially combine edge information for progressive upsampling and downsampling, ultimately outputting multiple results to the decoding network.

8. The method for detecting small targets in unmanned aerial vehicles based on edge information flow guidance according to claim 7, characterized in that, The info-PAN structure specifically includes: Edge information is extracted from the features of layer P2 through the information flow module to form edge information of layers P4, P3 and P5; In the upsampling branch, the input of P5 layer is fed into P4 layer via a convolutional block; in P4 layer, the output of P5 layer, the input of P4 layer, and the edge information of P4 layer are coupled through the SFB module and output to P3 layer; in P3 layer, the output of P4 layer, the input of P3 layer, and the edge information of P3 layer are coupled through the SFB module and output to the downsampling branch. In the downsampling branch, the output of the upsampling branch of layer P3 is received in layer P3 and processed by a convolutional block before being input to the decoding network and layer P4. In layer P4, the output of the downsampling branch of layer P3, the output of the upsampling branch of layer P4, and the edge information of layer P4 are coupled through the SFB module and output to layer P5 and the decoding network. In layer P5, the output of the downsampling branch of layer P4, the output of the upsampling branch of layer P5, and the edge information of layer P5 are coupled through the SFB module and output to the decoding network.

9. A method for detecting small targets on a UAV based on edge information flow guidance according to claim 7, characterized in that, The information flow module includes a GE module and three convolutional blocks connected in sequence. The GE module receives the input of the P2 layer, and the three convolutional blocks output the edge information of the P4, P3 and P5 layers respectively.

Images