A method, system and device for detecting small targets in a remote sensing image of a UAV based on an MSTY network

Abstract

A kind of unmanned aerial vehicle remote sensing image small target detection method, system and equipment based on MSTY network, its method includes: based on YOLOv11 architecture, constructs unmanned aerial vehicle remote sensing image small target detection MSTY network model;Wherein, local-global attention module is introduced in backbone network, for the modeling of local texture details and global context association;Cross-scale label aggregation module is introduced in neck network, for the adaptive fusion of multi-scale features in the process of cross-layer feature propagation, and the corresponding number of output features is obtained by multi-head attention mechanism;Scale-aware anchor-free detection head is introduced in detection head, respectively using each output feature, adaptive positioning and classification for different scale targets are realized;System and equipment are used to realize the method;The present application is directed to the detection enhancement scheme of small target and multi-scale scene, is improved from four levels of feature extraction, multi-scale fusion, detection head optimization and loss function, effectively improves the target detection performance in complex scene.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, and further relates to deep learning and digital image processing technology, specifically a method, system and device for small target detection in UAV remote sensing images based on MSTY network. Background Technology

[0002] With the rapid development of UAV technology and remote sensing imaging technology, UAV remote sensing images are widely used in fields such as refined urban management, environmental monitoring, emergency response, and target reconnaissance. Compared with traditional aerial and satellite remote sensing, UAVs have advantages such as low flight altitude, high imaging resolution, and flexible deployment. However, this also results in generally smaller target scales in the images, especially in large-scale urban scenes, where small targets often occupy only a few pixels and are easily submerged in complex backgrounds. Therefore, small target detection in UAV remote sensing images has become one of the research hotspots in the fields of computer vision and remote sensing information processing.

[0003] For the problem of small target detection in UAV remote sensing images, domestic and international scholars have mainly conducted research based on deep learning target detection frameworks. Existing methods can be roughly divided into two-stage detection methods and single-stage detection methods. The former, represented by Faster R-CNN, relies on candidate region generation and fine-grained classification regression, and has high detection accuracy; the latter, represented by YOLO, SSD, etc., emphasizes end-to-end learning and real-time performance. To enhance the feature representation ability of small targets, related research has introduced techniques such as feature pyramid networks, multi-scale feature fusion, attention mechanisms, and contextual information modeling, which have improved the detection effect of small targets in UAV remote sensing images to a certain extent.

[0004] While existing methods have achieved good detection performance on standard datasets, they still face many challenges in complex real-world scenarios. On the one hand, small targets are prone to feature weakening and information loss in deep networks, leading to limited detection accuracy. On the other hand, UAV remote sensing images typically have complex backgrounds, dense target distribution, and uneven class distribution, increasing the risk of false positives and false negatives. Furthermore, some improved methods rely on complex network structures, resulting in high computational costs and making it difficult to balance detection accuracy and real-time performance. Therefore, exploring more efficient, robust, and applicable methods for small target detection in UAV remote sensing images in complex scenarios remains a significant research endeavor.

[0005] Patent application CN120894537A discloses a method for small target detection in UAV images by fusing spatial and frequency features. The method involves several steps: first, inputting image data into a small target feature preservation network, where spatial and frequency features are extracted by spatial and frequency feature extraction branches, respectively; then, fusing the spatial and frequency features, encoding the fused feature map, and then performing cross-scale feature fusion to obtain an image feature sequence; finally, sending the filtered features to a decoder, where a decoder with an auxiliary prediction head completes target category and location detection. However, this invention still has shortcomings: the method has a long training time and relatively high computational complexity.

[0006] Patent application CN118212553A discloses a method for small target detection in UAV aerial photography based on an improved YOLOv7 algorithm. The method's implementation steps include: converting the UAV aerial photography dataset to a new format; dividing the converted dataset proportionally; improving the backbone network by constructing a coordinate convolution stacked branch and concatenating the features extracted from the backbone network with the features from the fusion layer; improving the Neck by constructing a deep routing attention module and passing the fused features through a feature fusion strategy module; improving the loss function by proposing a hybrid loss function as the localization loss in the prediction head; setting training parameters and saving weight data; selecting the optimal weights from the validation set and loading them into the network; and testing with a test set. However, this invention still has shortcomings, as the target detection accuracy remains relatively low.

[0007] Defects and shortcomings of existing technology: 1. Traditional algorithms rely on relatively fixed feature templates and hand-designed feature descriptors. This makes them unable to effectively capture the unique local information of small targets, leading to the easy overlooking or misjudgment of small targets. For example, when an object is small in scale within an image, its edges and details are often lost, and background noise or interference from other objects increases the difficulty of detection. Furthermore, because small targets occupy a relatively small visual area, the algorithm may not have sufficient contextual information to determine their presence, further exacerbating the probability of missed detections.

[0008] 2. While existing object detection models have achieved significant results in many application scenarios, they still face challenges in feature extraction and accurate location localization of small objects. Particularly in multi-scale fusion, models often struggle to capture feature information at different scales simultaneously, leading to small objects being overlooked or falsely detected in complex backgrounds. The inadequacy of multi-scale feature fusion reduces the model's ability to preserve details, especially in low-contrast and blurred edge conditions, significantly impacting object detection accuracy. Furthermore, existing models also have issues with the accuracy of small object localization, particularly when the features of the small object are too similar to the background, making it difficult for the detection network to achieve accurate boundary regression, further affecting detection accuracy. Summary of the Invention

[0009] To overcome the shortcomings of the prior art, the present invention aims to provide a method, system, and device for small target detection in UAV remote sensing images based on the MSTY network. By introducing a Local-Global Attention (LGAB) module into the backbone network, depthwise separable convolution and window self-attention can be combined to achieve collaborative modeling of local texture details and global contextual dependencies. By designing a cross-scale label aggregation (CTA) module in the feature fusion stage, multi-scale features can be adaptively fused at the label level, ensuring the effective preservation of high-resolution detail information during cross-layer propagation. By constructing a scale-aware anchorless detection head (SAAH) in the detection head, adaptive localization and classification of targets at different scales can be achieved. By introducing a hybrid loss function combining area weighting and scale penalty, the optimization contribution of small targets during training can be enhanced, significantly improving the accuracy and robustness of small target detection in UAV remote sensing images while maintaining the computational efficiency of YOLOv11.

[0010] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for small target detection in UAV remote sensing images based on MSTY network includes the following steps: Step 1: Based on the YOLOv11 architecture, construct the MSTY network model for small target detection in UAV remote sensing images. Specifically, a local-global attention module is introduced into the backbone network for modeling the relationship between local texture details and global context; a cross-scale label aggregation module is introduced into the neck network for adaptive fusion of multi-scale features at the label level during cross-layer feature propagation, and a corresponding number of output features are obtained through a multi-head attention mechanism; a scale-aware, anchor-free detection head is introduced into the detection head, utilizing each output feature to achieve adaptive localization and classification of targets at different scales. Step 2: Train the model using the UAV remote sensing image dataset, and use the trained model to detect small targets in UAV remote sensing images.

[0011] The backbone network includes several feature extractors, and there are multiple local-global attention modules, which are introduced into the output of one of the feature extractors in the backbone network. Except for the first level, the output of each local-global attention module is divided into two paths. One path is sent to the next level feature extractor after passing through a convolution module to continue extracting features, and the other path is sent to the neck network to participate in feature aggregation. The output of the last level feature extractor of the backbone network is enhanced by pyramid pooling, and then sent to the neck network after passing through a cross-stage local spatial attention mechanism.

[0012] The local-global attention module includes a depthwise separable convolution (DWConv), a batch normalized (BN) function, a layer normalized (LN) function, two GELU activation functions, a 1x1 convolution (Conv1), a multi-head self-attention mechanism (MHSA), a compression and activation mechanism (SE), and a dot product. Summing two ;in: Given input features The global features are obtained by sequentially applying depthwise separable convolution (DWConv), batch normalization (BN), the first GELU activation function, and a 1x1 convolution (Conv1). Input features Inputting the multi-head self-attention mechanism MHSA yields local features. Use summation global features and local features By fusing, features are obtained. ; Features Features are obtained by sequentially applying layer-normalized LN and a second GELU activation function. ; Features Input compression and excitation mechanism SE obtains features Use dot product Features and characteristics By fusing, features are obtained. Finally, use summation. Features and input features The features are fused to obtain the output features. ; The formula for the Local-Global Attention Module (LGAB) is expressed as follows: in, express convolution, Indicates batch normalization, This represents depthwise separable convolution; This represents the GELU activation function; This indicates compression and incentive mechanisms.

[0013] The cross-scale label aggregation module is introduced between the output of the neck network and the input of the detection head. It fuses features from different levels of the neck network through a multi-head attention mechanism and outputs several output features equal to the number of levels.

[0014] The cross-scale label aggregation module (CTA) includes six 1x1 convolutional modules (Conv1), three Flatten modules, and three multi-head self-attention mechanisms (MHSA); wherein: Given multi-scale features at three different levels of the neck network , and ,feature The features are obtained by sequentially passing the first 1x1 convolution Conv1 and the first Flatten module. ;feature Features are obtained by sequentially passing the second 1x1 convolution Conv1 and the second Flatten module. ;feature The features are obtained by sequentially passing the third 1*1 convolution Conv1 and the third Flatten module. ; Features ,feature and characteristics Inputting the first multi-head self-attention mechanism MHSA, we obtain the features. ; Features ,feature and characteristics Inputting a second multi-head self-attention mechanism (MHSA) yields features. ; Features ,feature and characteristics Inputting a third multi-head self-attention mechanism (MHSA) yields features. ; Features Input the fourth 1x1 convolution Conv1 to obtain the output features. ; Features Input the fifth 1x1 convolution Conv1 to obtain the output features. ; Features Input the sixth 1x1 convolution Conv1 to obtain the output features. ; The formula for the cross-scale marker aggregation module (CTA) is expressed as follows: The number of scale-aware anchorless detection heads is the same as the output features of the neck network. Each scale-aware anchorless detection head takes one output feature as input to perform target detection at the corresponding scale. The fusion result of the scale-aware anchorless detection heads is used as the final detection result.

[0015] The scale-aware, anchorless detection head SAAH includes a global pooling module GAP, four 3x3 convolutional modules Conv3, three 1x1 convolutional modules Conv1, and a dot product. ;in: Given the output features from the cross-scale label aggregation module CTA , and One of the features is defined as feature The features are obtained by sequentially passing them through a global pooling module (GAP) and a first 3x3 convolution (Conv3). Use dot product Features and characteristics By fusing, features are obtained. ;feature The regression features are obtained by sequentially passing the second 3x3 convolution Conv3 and the first 1x1 convolution Conv1. ;feature The classification features are obtained by sequentially performing a third 3x3 convolution (Conv3) and a second 1x1 convolution (Conv1). ;feature The IoU features are obtained by sequentially performing a fourth 3x3 convolution (Conv3) and a third 1x1 convolution (Conv1). ; The SAAH formula for a scale-aware, anchorless detection head is expressed as follows: .

[0016] In step 2, during the training phase, a hybrid loss function combining area weighting and scale penalty terms is introduced to optimize classification, regression, and IoU in the cross-scale label aggregation module (CTA), thereby enhancing the recognition and localization accuracy of small targets during training. The formula for the hybrid loss function is as follows: in, , and These are weighting coefficients. This represents the predicted probability of the correct class. It is the focusing parameter. This represents the intersection-union ratio between the predicted bounding box and the ground truth bounding box. It is the Euclidean distance between their centers. This represents the diagonal length of the smallest bounding box. Measuring aspect ratio consistency It is a balancing factor. This represents the predicted IoU perceived score. It is the modulation factor. Scale-aware weights.

[0017] This invention also provides a small target detection system for UAV remote sensing images based on MSTY network, comprising: The network construction module is used to build a small target detection MSTY network model for UAV remote sensing images based on the YOLOv11 architecture. Specifically, a local-global attention module is introduced into the backbone network for modeling the relationship between local texture details and global context; a cross-scale label aggregation module is introduced into the neck network for adaptive fusion of multi-scale features at the label level during cross-layer feature propagation, and a corresponding number of output features are obtained through a multi-head attention mechanism; a scale-aware, anchor-free detection head is introduced into the detection head, which utilizes each output feature to achieve adaptive localization and classification of targets at different scales. The network training and target detection module is used to train the model using a UAV remote sensing image dataset and to use the trained model to detect small targets in UAV remote sensing images.

[0018] This invention also provides a small target detection device for UAV remote sensing images based on MSTY network, comprising: Memory: A computer program that stores the above-mentioned method for detecting small targets in UAV remote sensing images based on MSTY network, and is a computer-readable device; Processor: Used to implement the method for small target detection in UAV remote sensing images based on MSTY network when executing the computer program.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention constructs an MSTY network based on YOLOv11. By introducing a local-global attention module into the backbone network, it effectively fuses the detailed features of small targets with global semantics, thereby improving feature representation capabilities. A cross-scale label aggregation module is employed in the neck network to enhance multi-scale feature interaction and improve adaptability to targets with varying scales. A scale-aware, anchor-free bounding box mechanism is introduced into the detection head to reduce reliance on prior bounding boxes, achieving more accurate localization and classification. Therefore, this invention can significantly improve the accuracy and robustness of small target detection in UAV remote sensing images.

[0020] 2. This invention proposes a Local-Global Attention Module (LGAB). This module first extracts texture features with strong local perceptual capabilities through Depthily Separable Convolution (DWConv), batch normalization, and GELU activation, and then performs feature mapping via 1*1 convolution (Conv1*1). Simultaneously, a Multi-Head Self-Attention (MHSA) mechanism is introduced to model the global dependencies of the input features, thereby obtaining both local and global features. Subsequently, element-wise summation is used to achieve dual-branch feature fusion, and batch normalization and nonlinear activation are combined to further enhance expressive power. Based on this, a Sequential Energization (SE) compression activation mechanism is introduced for channel recalibration, and key response regions are strengthened through dot product operations. Finally, the output is obtained by fusing the residual connections with the input features. This module achieves deep coupling between local detail enhancement and global semantic modeling within a single structure, significantly improving the discrimination ability of small targets in complex backgrounds, demonstrating superior fusion modeling advantages over traditional convolutional or single attention structures.

[0021] 3. This invention designs a cross-scale label aggregation module (CTA) for efficient interaction and fine-grained fusion of multi-scale features in neck networks. This module performs channel alignment for features at different levels using 1x1 convolutions (Conv1x1), and converts spatial features into serialized labels using the Flatten operation. Then, a multi-head self-attention mechanism (MHSA) is used to model cross-scale dependencies at the label level, achieving global feature interaction and semantic alignment. Features at each scale are enhanced by attention and then their spatial structure is restored using 1x1 convolutions, forming a multi-branch parallel cross-scale fusion path. This method overcomes the limitation of traditional FPN / PAN structures that rely solely on layer-by-layer fusion, achieving deep collaborative modeling and significantly enhancing the model's consistency in representing multiple scales, especially small targets.

[0022] 4. To further enhance the adaptability of the detection head to scale changes, this invention constructs a scale-aware, anchor-free detection head, SAAH. This detection head abandons the traditional anchor-based design paradigm and introduces a global pooling module (GAP), enabling the model to adaptively adjust the prediction strategy for different spatial locations based on feature responses. During classification and localization, it achieves effective alignment between classification confidence and localization accuracy, thereby reducing false positives and false negatives caused by scale mismatch. Compared to traditional detection heads, this method exhibits stronger generalization ability and localization accuracy in complex scale distribution scenarios, and is particularly suitable for detecting small targets with drastic size changes in UAV remote sensing images.

[0023] 5. Regarding training optimization, this invention proposes a hybrid loss function combining area weighting and scale penalty terms, providing unified modeling across classification loss, regression loss, and IoU optimization. This loss function introduces weighting coefficients and focusing parameters based on Focal Loss to enhance the learning weights of difficult and small target samples. In the localization loss, it achieves fine-grained optimization of the predicted box geometry by introducing Euclidean distance of the center point, minimum bounding box diagonal length, and aspect ratio consistency constraints. Simultaneously, it designs IoU-aware scores and modulation factors to improve the matching accuracy between the predicted results and the ground truth boxes. Furthermore, by introducing scale-aware weights, it applies differentiated constraints to targets of different scales, significantly enhancing the regression stability and localization accuracy of small targets. This hybrid loss function optimizes the training process from multiple dimensions, demonstrating significant advantages in both accuracy and robustness compared to traditional single-loss designs.

[0024] In summary, this invention provides a detection enhancement scheme for small targets and multi-scale scenes, making systematic improvements from four aspects: feature extraction, multi-scale fusion, detection head optimization, and loss function design, effectively enhancing target detection performance in complex scenes. Attached Figure Description

[0025] Figure 1 This is a flowchart of the implementation method of the present invention.

[0026] Figure 2 This is a schematic diagram of the overall network structure of the present invention.

[0027] Figure 3 This is a schematic diagram of the Local-Global Attention Block (LGAB) structure of the present invention.

[0028] Figure 4 This is a schematic diagram of the Cross-Scale Token Aggregation (CTA) module of the present invention.

[0029] Figure 5This is a schematic diagram of the scale-aware anchor-free head (SAAH) structure of the present invention.

[0030] Figure 6 This is a comparison chart of the small target detection accuracy of the present invention and existing methods. Detailed Implementation

[0031] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings.

[0032] like Figure 1 As shown, a method for small target detection in UAV remote sensing images based on MSTY networks includes the following steps: Step 1: As Figure 2 As shown, an MSTY network model for small target detection in UAV remote sensing images is constructed. The MSTY network model for small target detection in UAV remote sensing images in step 1 is built on the YOLOv11 architecture. A Local-Global Attention Block (LGAB) module is introduced in the backbone network to model the relationship between local texture details and global context, improving the ability to recognize small target features. A Cross-Scale Token Aggregation (CTA) module is introduced in the neck network to adaptively fuse multi-scale features at the tag level during cross-layer feature propagation and obtain a corresponding number of output features through a multi-head attention mechanism. In the head part, a Scale-Aware Anchor-Free Head (SAAH) is introduced to utilize each output feature to achieve adaptive localization and classification of targets at different scales. The backbone network includes an input module, five Conv,k=3 modules, four C3k2 modules, three local-global attention modules (LGAB), one C2 position-sensitive attention module (C2PSA), and one fast spatial pyramid pooling module (SPPF). The Local-Global Attention Module (LGAB) is sequentially introduced between the first C3k2 module and the third Conv,k=3 module, between the second C3k2 module and the fourth Conv,k=3 module, and between the third C3k2 module and the fifth Conv,k=3 module in the backbone network. Specifically, the output of the second local-global attention module LGAB is fused with the first concat operation of the neck network, and the output of the third local-global attention module is fused with the second concat operation of the neck network.

[0033] Step 101: As Figure 3 As shown, a Local-Global Attention Module (LGAB) is constructed. The LGAB includes a depthwise separable convolution (DWConv), a batch normalized normalization (BN), a layer normalized normalization (LN), two GELU activation functions, a 1x1 convolution (Conv1), a multi-head self-attention mechanism (MHSA), a compression and activation mechanism (SE), and a dot product. Summing two Specifically: Given input features Input features The global features are obtained by sequentially inputting a depthwise separable convolution (DWConv), batch normalization (BN), the first GELU activation function, and a 1x1 convolution (Conv1). Input features Inputting the multi-head self-attention mechanism MHSA yields local features. Use summation global features and local features By fusing, features are obtained. ; Features Features are obtained by sequentially applying layer-normalized LN and a second GELU activation function. ; Features Input compression and excitation mechanism SE obtains features Use dot product Features and characteristics By fusing, features are obtained. Finally, use summation. Features and input features The features are fused to obtain the output features. ; The formula for the Local-Global Attention Module (LGAB) is expressed as follows: in, express convolution, Indicates batch normalization, This represents depthwise separable convolution; This represents the GELU activation function; This indicates compression and incentive mechanisms.

[0034] The Local-Global Attention (LGAB) module effectively fuses local texture information with global semantic information by combining depthwise separable convolution (DWConv) with multi-head self-attention (MHSA) mechanism. The convolutional branch enhances the detailed representation of small targets, while the self-attention branch strengthens the modeling ability for long-distance dependencies and contextual information. Furthermore, feature fusion further improves the discriminativeness and stability of feature representation. Overall, this module significantly enhances the model's ability to recognize small targets in complex backgrounds, while balancing computational efficiency and robustness.

[0035] Among them, the Multi-Head Self-Attention (MHSA) mechanism achieves multi-scale contextual reasoning by dynamically aggregating information across feature levels. Ultimately, the reconstructed features retain both local details from low-level features and abstract semantic representations from high-level features, thus forming a unified and semantically rich multi-scale feature representation.

[0036] Depthwise separable convolution (DWConv) decomposes standard convolution into channel-wise depthwise convolution and cross-channel pointwise convolution, significantly reducing the number of parameters and computational complexity while maintaining feature representation capabilities. Depthwise convolution extracts spatial features, while pointwise convolution achieves information fusion between channels, thus balancing efficiency and performance in lightweight model design, making it suitable for real-time visual tasks.

[0037] The cross-scale label aggregation module is introduced between the output of the neck network and the input of the detection head. It fuses features from different levels of the neck network through a multi-head attention mechanism and outputs several output features equal to the number of levels.

[0038] Step 102: As Figure 4 As shown, a cross-scale label aggregation module (CTA) is constructed, comprising six 1*1 convolutional modules (Conv1), three Flatten modules, and three multi-head self-attention mechanisms (MHSA); specifically: Given multi-scale features at three different levels of the neck network , and ,feature The features are obtained by sequentially passing the first 1x1 convolution Conv1 and the first Flatten module. ;feature Features are obtained by sequentially passing the second 1x1 convolution Conv1 and the second Flatten module. ;feature The features are obtained by sequentially passing the third 1*1 convolution Conv1 and the third Flatten module. ; Features ,feature and characteristics Inputting the first multi-head self-attention mechanism MHSA, we obtain the features. ; Features ,feature and characteristics Inputting a second multi-head self-attention mechanism (MHSA) yields features. ; Features ,feature and characteristics Inputting a third multi-head self-attention mechanism (MHSA) yields features. ; Features Input the fourth 1x1 convolution Conv1 to obtain the output features. ; Features Input the fifth 1x1 convolution Conv1 to obtain the output features. ; Features Input the sixth 1x1 convolution Conv1 to obtain the output features. ; The formula for the cross-scale marker aggregation module (CTA) is expressed as follows: The Cross-Scale Tag Aggregation (CTA) module aggregates multi-scale features. A tokenization process is performed, and a multi-head self-attention mechanism (MHSA) is introduced at the tokenization level to achieve global information interaction and adaptive fusion between features of different resolutions. This module uniformly maps features at each scale to the tokenization space, enhancing cross-scale dependency modeling capabilities while maintaining the independent expression of semantic information at each layer, thereby effectively alleviating the information fragmentation problem. Overall, CTA can improve the collaborative expression capability of multi-scale features and enhance the model's perception and detection performance for targets at different scales, especially showing better adaptability to small targets with significant scale changes.

[0039] The number of scale-aware anchorless detection heads is the same as the output features of the neck network. Each scale-aware anchorless detection head takes one output feature as input to perform target detection at the corresponding scale. The fusion result of the scale-aware anchorless detection heads is used as the final detection result.

[0040] Step 103: As Figure 5 As shown, a scale-aware, anchorless detection head, SAAH, is constructed, consisting of a global pooling module (GAP), four 3x3 convolutional modules (Conv3), three 1x1 convolutional modules (Conv1), and a dot product. Specifically: Given the output features from the cross-scale label aggregation module CTA , and One of the features is defined as feature The features are obtained by sequentially inputting a global pooling module (GAP) and a first 3x3 convolutional module (Conv3). Use dot product Features and characteristics By fusing, features are obtained. ; Features By sequentially inputting the second 3x3 convolution Conv3 and the first 1x1 convolution Conv1, the regression features are obtained. ; Features By sequentially inputting the third 3x3 convolution Conv3 and the second 1x1 convolution Conv1, the classification features are obtained. ; Features By sequentially inputting the fourth 3x3 convolution Conv3 and the third 1x1 convolution Conv1, we obtain the IoU features. ; The SAAH formula for a scale-aware, anchorless detection head is expressed as follows: This detection head is a multi-scale, anchorless detection head, corresponding to three-layer fusion features. , representing high, medium, and low resolution features, respectively. This module takes the fused features output by the cross-scale label aggregation module (CTA) as input. By introducing global pooling and mechanisms, it achieves adaptive perception and enhancement of features at different scales. Based on cross-scale fused features, this module utilizes global contextual information to fuse features, thereby highlighting key region responses and suppressing redundant information; subsequently, convolutional branches are used to generate classification, regression, and IoU prediction features, improving the representational independence and learning efficiency between tasks. SAAH can effectively improve the localization and classification performance of multi-scale targets within an anchor-free framework, especially enhancing the detection capability of targets with significant scale changes.

[0041] Step 2: Set the number of training rounds to ≥300, the batch size to ≥64, and train the MSTY network model for small target detection in UAV remote sensing images using the training set in the publicly available dataset for UAV remote sensing image small target detection. Output the training weight file MSTY.pt. During the training phase, a hybrid loss function combining area weighting and scale penalty terms is introduced to optimize classification, regression, and IoU in the cross-scale label aggregation module (CTA), enhancing the recognition and localization accuracy of small targets during training. The formula for the hybrid loss function is as follows: in, , and These are weighting coefficients; This represents the predicted probability of the correct class. It is the focusing parameter; This represents the intersection-union ratio between the predicted bounding box and the ground truth bounding box; It is the Euclidean distance between their centers; This represents the diagonal length of the smallest bounding box; Measuring aspect ratio consistency; It is a balancing factor; This represents the predicted IoU perceived score. It is the modulation factor. Scale-aware weights.

[0042] Step 3: With a batch size ≥ 32, use the test set from the public dataset for small target detection in UAV remote sensing images and the weight file MSTY.pt obtained in Step 2 to perform small target detection on the MSTY network model for small target detection in UAV remote sensing images, and obtain the small target detection results.

[0043] This invention also provides a small target detection system for UAV remote sensing images based on MSTY network, comprising: The network construction module is used to implement the MSTY network model for small target detection in UAV remote sensing images based on the YOLOv11 architecture in step 1. Specifically, a local-global attention module is introduced into the backbone network for modeling the relationship between local texture details and global context; a cross-scale label aggregation module is introduced into the neck network for adaptive fusion of multi-scale features at the label level during cross-layer feature propagation, and a corresponding number of output features are obtained through a multi-head attention mechanism; and a scale-aware, anchor-free detection head is introduced into the detection head to achieve adaptive localization and classification of targets at different scales by utilizing each output feature. The network training module is used to train the MSTY network model for small target detection in UAV remote sensing images using the training set in the publicly available dataset for small target detection in UAV remote sensing images in step 2, and output the training weight file. The target detection module is used to perform small target detection on the MSTY network model of UAV remote sensing image small target detection in step 3 using the test set in the public dataset of UAV remote sensing image small target detection and the training weight file, and obtain the small target detection result.

[0044] This invention also provides a small target detection device for UAV remote sensing images based on MSTY network, comprising: Memory: A computer program that stores the above-mentioned method for detecting small targets in UAV remote sensing images based on MSTY network, and is a computer-readable device; Processor: Used to implement the method for small target detection in UAV remote sensing images based on MSTY network when executing the computer program.

[0045] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, enables the implementation of the aforementioned method for small target detection in UAV remote sensing images based on an MSTY network.

[0046] The effects of the present invention will be further illustrated below with simulation experiments: 1. Simulation Experiment Conditions The hardware platform for the simulation experiment of this invention is: Intel i9-13900K processor with a main frequency of 3.0GHz and 64G RAM.

[0047] The software platform of the simulation experiment platform of this invention is: Windows 11 operating system and PyCharm, PyTorch 2.1.0, CUDA 12.1.

[0048] 2. Simulation Steps The training set from the publicly available dataset for small target detection in UAV remote sensing images is input into the UAV remote sensing image small target detection network based on the MSTY network for training. The training process is as follows: the input image is processed by a feature extraction network to extract features, resulting in feature maps of different scales. These feature maps are then reconstructed to obtain more refined feature maps. Based on this, the loss is calculated to complete the small target detection based on this invention. The number of iterations (epochs) is set to 300.

[0049] 3. Simulation content and result analysis The simulation experiment of this invention uses this invention to perform small target detection processing on remote sensing images of 50 UAVs, and the results are as follows: Figure 6As shown, the results are averaged.

[0050] The following is combined with Figure 6 The simulation effects of the present invention will be further described.

[0051] Figure 6 This is a comparison chart of the road extraction accuracy of the present invention and existing methods. The comparison algorithm is EF-DETR (Cheng S, Song J, Zhou M, et al. EF-DETR: A Lightweight Transformer-Based ObjectDetector With an Encoder-Free Neck[J].IEEE transactions on industrialinformatics, 2024(11):20.), and the evaluation metric is recall (R), which measures the model's ability to correctly identify all correct samples.

[0052] like Figure 6 As shown in the experimental results, the proposed method for small target detection in UAV remote sensing images based on MSTY network has a higher R value than EF-DETR, which fully demonstrates the effectiveness of the method in meeting the requirements of high-precision small target detection.

Claims

1. A method for small target detection in UAV remote sensing images based on MSTY network, characterized in that, Includes the following steps: Step 1: Based on the YOLOv11 architecture, construct the MSTY network model for small target detection in UAV remote sensing images. Specifically, a local-global attention module is introduced into the backbone network for modeling the relationship between local texture details and global context; a cross-scale label aggregation module is introduced into the neck network for adaptive fusion of multi-scale features at the label level during cross-layer feature propagation, and a corresponding number of output features are obtained through a multi-head attention mechanism; a scale-aware, anchor-free detection head is introduced into the detection head, utilizing each output feature to achieve adaptive localization and classification of targets at different scales. Step 2: Train the model using the UAV remote sensing image dataset, and use the trained model to detect small targets in UAV remote sensing images.

2. The method for small target detection in UAV remote sensing images based on MSTY network according to claim 1, characterized in that, The backbone network includes several feature extractors, and there are multiple local-global attention modules, which are introduced into the output of one of the feature extractors in the backbone network. Except for the first level, the output of each local-global attention module is divided into two paths. One path is sent to the next level feature extractor after passing through a convolution module to continue extracting features, and the other path is sent to the neck network to participate in feature aggregation. The output of the last level feature extractor of the backbone network is enhanced by pyramid pooling, and then sent to the neck network after passing through a cross-stage local spatial attention mechanism.

3. A method for small target detection in UAV remote sensing images based on an MSTY network according to claim 1 or 2, characterized in that, The local-global attention module includes a depthwise separable convolution (DWConv), a batch normalized (BN) function, a layer normalized (LN) function, two GELU activation functions, a 1x1 convolution (Conv1), a multi-head self-attention mechanism (MHSA), a compression and activation mechanism (SE), and a dot product. Summing two ;in: Given input features The global features are obtained by sequentially applying depthwise separable convolution (DWConv), batch normalization (BN), the first GELU activation function, and a 1x1 convolution (Conv1). Input features Inputting the multi-head self-attention mechanism MHSA yields local features. Use summation global features and local features By fusing, features are obtained. ; Features Features are obtained by sequentially applying layer-normalized LN and a second GELU activation function. ; Features Input compression and excitation mechanism SE obtains features Use dot product Features and characteristics By fusing, features are obtained. Finally, use summation. Features and input features The features are fused to obtain the output features. .

4. The method for small target detection in UAV remote sensing images based on MSTY network according to claim 1, characterized in that, The cross-scale label aggregation module is introduced between the output of the neck network and the input of the detection head. It fuses features from different levels of the neck network through a multi-head attention mechanism and outputs several output features equal to the number of levels.

5. A method for small target detection in UAV remote sensing images based on an MSTY network according to claim 1 or 4, characterized in that, The cross-scale label aggregation module (CTA) includes six 1x1 convolutional modules (Conv1), three Flatten modules, and three multi-head self-attention mechanisms (MHSA); wherein: Given multi-scale features at three different levels of the neck network , and ,feature The features are obtained by sequentially passing the first 1x1 convolution Conv1 and the first Flatten module. ;feature Features are obtained by sequentially passing the second 1x1 convolution Conv1 and the second Flatten module. ;feature The features are obtained by sequentially passing the third 1*1 convolution Conv1 and the third Flatten module. ; Features ,feature and characteristics Inputting the first multi-head self-attention mechanism MHSA, we obtain the features. ; Features ,feature and characteristics Inputting a second multi-head self-attention mechanism (MHSA) yields features. ; Features ,feature and characteristics Inputting a third multi-head self-attention mechanism (MHSA) yields features. ; Features Input the fourth 1x1 convolution Conv1 to obtain the output features. ; Features Input the fifth 1x1 convolution Conv1 to obtain the output features. ; Features Input the sixth 1x1 convolution Conv1 to obtain the output features. .

6. The method for small target detection in UAV remote sensing images based on MSTY network according to claim 1, characterized in that, The number of scale-aware anchorless detection heads is the same as the output features of the neck network. Each scale-aware anchorless detection head takes one output feature as input to perform target detection at the corresponding scale. The fusion result of the scale-aware anchorless detection heads is used as the final detection result.

7. A method for small target detection in UAV remote sensing images based on an MSTY network according to claim 1 or 6, characterized in that, The scale-aware, anchorless detection head SAAH includes a global pooling module GAP, four 3x3 convolutional modules Conv3, three 1x1 convolutional modules Conv1, and a dot product. ;in: Given the output features from the cross-scale label aggregation module CTA , and One of the features is defined as feature The features are obtained by sequentially passing them through a global pooling module (GAP) and a first 3x3 convolution (Conv3). Use dot product Features and characteristics By fusing, features are obtained. ;feature The regression features are obtained by sequentially passing the second 3x3 convolution Conv3 and the first 1x1 convolution Conv1. ;feature The classification features are obtained by sequentially performing a third 3x3 convolution (Conv3) and a second 1x1 convolution (Conv1). ;feature The IoU features are obtained by sequentially performing a fourth 3x3 convolution (Conv3) and a third 1x1 convolution (Conv1). .

8. The method for small target detection in UAV remote sensing images based on MSTY network according to claim 1, characterized in that, In step 2, during the training phase, a hybrid loss function combining area weighting and scale penalty terms is introduced to optimize classification, regression, and IoU in the cross-scale label aggregation module (CTA), thereby enhancing the recognition and localization accuracy of small targets during training. The formula for the hybrid loss function is as follows: in, , and These are weighting coefficients. This represents the predicted probability of the correct class. It is the focusing parameter. This represents the intersection-union ratio between the predicted bounding box and the ground truth bounding box. It is the Euclidean distance between their centers. This represents the diagonal length of the smallest bounding box. Measuring aspect ratio consistency It is a balancing factor. This represents the predicted IoU perceived score. It is the modulation factor. Scale-aware weights.

9. A small target detection system for UAV remote sensing images based on an MSTY network, used to implement the method described in claim 1, characterized in that, include: The network construction module is used to build a small target detection MSTY network model for UAV remote sensing images based on the YOLOv11 architecture. Specifically, a local-global attention module is introduced into the backbone network for modeling the relationship between local texture details and global context; a cross-scale label aggregation module is introduced into the neck network for adaptive fusion of multi-scale features at the label level during cross-layer feature propagation, and a corresponding number of output features are obtained through a multi-head attention mechanism; a scale-aware, anchor-free detection head is introduced into the detection head, which utilizes each output feature to achieve adaptive localization and classification of targets at different scales. The network training and target detection module is used to train the model using a UAV remote sensing image dataset and to use the trained model to detect small targets in UAV remote sensing images.

10. A small target detection device for UAV remote sensing images based on MSTY network, characterized in that, include: Memory: A computer program for a method for detecting small targets in UAV remote sensing images based on an MSTY network, as described in any one of claims 1-8, is a computer-readable device; Processor: Used to implement the method for small target detection in UAV remote sensing images based on MSTY network as described in any one of claims 1-8 when executing the computer program.

Images