Improved YOLOv10 network for detecting small target in remote sensing image

Abstract

The invention provides an improved YOLOv10 network for detecting a small target in a remote sensing image. The improved YOLOv10 network is used for enhancing the detection performance of the small target in the remote sensing image. Existing first-stage detection methods, such as YOLOv10, can realize rapid detection, but still face challenges in small target detection under a complex background. According to the method, the HATSC module is constructed by introducing the mixed attention converter and the space and channel reconstruction convolution module, so that the feature extraction capability of the model on local details under a complex background is effectively improved. Meanwhile, the normalized Wasserstein distance is used as a part of a loss function, so that the sensitivity to the position deviation of the small target is reduced. The HATSC module is integrated in the HATSC-YOLOv10 network, optimization is carried out in the stages of feature extraction, feature fusion and detection heads, the detection precision and efficiency of small targets in remote sensing images are improved, and the method is suitable for multiple fields such as environment monitoring, agricultural management and urban planning.

Description

Technical Field

[0001] The present invention relates to the field of algorithms, and in particular to an improved YOLOv10 network for small target detection in remote sensing images. Background Art

[0002] With the rapid development of remote sensing technology, remote sensing image data has shown explosive growth, which has put forward higher requirements on the analysis ability of remote sensing images. As one of the core tasks of remote sensing image analysis, remote sensing image target detection technology aims to automatically identify and extract specific ground objects such as buildings, vehicles, ships, roads, etc. from massive remote sensing images. However, traditional manual labeling and detection methods can no longer meet the needs of efficient and accurate target detection. Therefore, target detection methods based on deep learning have emerged and have achieved remarkable results in the field of remote sensing image target detection.

[0003] As a representative of single-stage target detection algorithms, the YOLO (You Only Look Once) series of algorithms have attracted widespread attention for their efficient and real-time detection performance. As the latest member of the YOLO series, YOLOv10 has achieved significant improvements in detection accuracy and speed. However, there are still some challenges for the detection of small targets in remote sensing images, such as insufficient feature extraction capabilities of local details under complex backgrounds and sensitivity of the loss function to small target position deviations. In order to solve these problems, the present invention proposes a new improved YOLOv10 network based on hybrid attention transformer (HAT) and spatial and channel reconstruction convolution (SCconv), namely HATSC-YOLOv10, to improve the detection performance of YOLOv10 in small targets in remote sensing images.

[0004] For the detection of small targets in remote sensing images, the existing technologies still have the following shortcomings: (1) Although the two-stage detection method can achieve high accuracy, it is computationally intensive and is not suitable for real-time applications.

[0005] (2) Although the one-stage detection method can achieve a high detection speed, its detection accuracy is relatively low and it may not be able to effectively detect small targets in the image.

[0006] (3) The existing one-stage detection method is not capable of extracting features of local details in complex backgrounds, resulting in limited performance in small target detection.

[0007] (4) The loss function is sensitive to the position deviation of small targets, making it difficult to improve the detection accuracy of small targets during network training.

[0008] In order to solve the above problems, the applicant proposed an improved YOLOv10 network for small target detection in remote sensing images. Summary of the invention

[0009] The purpose of the present invention is to provide an improved YOLOv10 network for small target detection in remote sensing images to solve the problems in the prior art.

[0010] To achieve the above object, the present invention provides the following technical solution: an improved YOLOv10 network for small target detection in remote sensing images, comprising: A transformer (SC) module that combines a hybrid attention transformer and spatial and channel reconstruction convolutions for deep feature extraction and fusion of small objects in remote sensing images to enhance the feature extraction capability of local details under complex backgrounds; A YOLOv10 network structure, in which the converter SC module is integrated to form a converter SC-YOLOv10 network structure to improve the detection performance of small objects in remote sensing images; and a loss function in which NWD is introduced to measure the similarity between the bounding boxes of detected small objects to reduce the sensitivity of the loss function to small target position deviations.

[0011] Optionally, the converter module is used to perform deep feature extraction and fusion on image details.

[0012] Optionally, the channel reconstruction convolution is used to compress the detection network model to improve the performance of the detection network.

[0013] Optionally, the NWD is used to measure the similarity between bounding boxes of small objects detected in the loss function, so that the loss function is insensitive to position deviations of small objects.

[0014] Optionally, the converter SC-YOLOv10 network structure improves the small target detection performance in remote sensing images by combining the converter SC module and the YOLOv10 network.

[0015] A method for improving the YOLOv10 network for small target detection in remote sensing images, characterized in that it comprises the following steps: a. Use the converter SC module to perform deep feature extraction and fusion of remote sensing images; b. Input the extracted features into the YOLOv10 network integrated with the converter SC module for target detection; c. Use the NWD-introduced loss function to optimize the position of the detected targets to reduce the sensitivity to small target position deviations.

[0016] Beneficial effects: Aiming at the shortcomings of the prior art, the purpose of the present invention is to propose a new improved YOLOv10 network based on hybrid attention transformer (HAT) and spatial and channel reconstruction convolution (SCconv), namely HATSC-YOLOv10, to improve the detection performance of YOLOv10 in small targets in remote sensing images. Specific performance includes: 1. Enhance the model's ability to extract features of local details in complex backgrounds.

[0017] 2. Reduce the sensitivity of the loss function to small target position deviations.

[0018] 3. Improve the detection performance of small targets in remote sensing images. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a schematic diagram of the flow chart of an embodiment of the present invention. DETAILED DESCRIPTION

[0020] The following describes the preferred embodiments of the present invention with reference to the drawings in the specification, so that the technical content is clearer and easier to understand. The present invention can be embodied in many different forms of embodiments, and the protection scope of the present invention is not limited to the embodiments mentioned in the text.

[0021] Example 1

[0022] Aiming at the challenge of small target detection in remote sensing images, the present invention proposes a new network architecture, HATSC-YOLOv10. Based on YOLOv10, the network integrates two key technologies, hybrid attention converter (HAT) and spatial and channel reconstruction convolution (SCconv), and introduces a new loss function measurement method (NWD) to improve the detection accuracy and efficiency of small targets in complex backgrounds. The technical solution of the present invention will be described in detail below.

[0023] Hybrid Attention Transformer (HAT) is an innovative feature extraction and fusion module designed to enhance the model's ability to focus on local details in complex backgrounds. HAT combines the advantages of channel attention and spatial attention, and extracts more refined and effective feature representations by adaptively adjusting the feature weights of different channels and spatial positions.

[0024] In remote sensing images, small targets are often submerged in complex backgrounds, and their features are relatively weak. When processing such images, traditional convolutional neural networks tend to ignore the features of small targets due to the interference of background noise. The HAT module introduces an attention mechanism that can dynamically adjust the weights of feature maps, allowing the model to pay more attention to small targets and the local details around them, thereby improving the detection accuracy of small targets.

[0025] In the HATSC-YOLOv10 network, HAT modules are embedded in different layers of the network to enhance attention to feature maps of different scales. Specifically, after each feature extraction stage, a HAT module is connected to perform dual adjustments of channel and spatial attention on the feature map. In this way, the network can more effectively extract and utilize small target features in remote sensing images, providing stronger support for subsequent classification and positioning tasks.

[0026] Spatial and Channel Reconstruction Convolution (SCconv) is a lightweight convolution operation that aims to maintain or even improve the performance of the model by reducing model parameters and computation. SCconv reconstructs and recombine the spatial and channel dimensions of the feature map, thereby achieving efficient use of features and compression of the model.

[0027] In the task of remote sensing image target detection, the complexity and computation of the model are one of the key factors affecting real-time performance. When processing high-resolution remote sensing images, traditional convolution operations often lead to too many model parameters and too much computation, which limits the real-time application capability of the model. The SCconv module, by introducing the idea of ​​spatial and channel reconstruction, can effectively reduce the complexity and computation of the model without sacrificing too much accuracy, thereby improving the real-time performance of the model.

[0028] In the HATSC-YOLOv10 network, the SCconv module is used to replace some traditional convolution operations to achieve network compression and acceleration. Specifically, in the feature extraction stage of the network, some convolutional layers are replaced by SCconv layers to reduce the parameters and computational complexity of the model. At the same time, in order to ensure that the performance of the model is not greatly affected, the SCconv layer fully considers the spatial and channel dimension information of the feature map during design and implementation to ensure the effective use and transmission of features.

[0029] Traditional target detection loss functions are usually measured based on parameters such as bounding box coordinates, aspect ratio, and classification probability. However, in the task of detecting small targets in remote sensing images, due to the small size and weak features of small targets, traditional loss functions are often sensitive to the position deviation of small targets, which can easily lead to difficulties in improving the detection accuracy of small targets during the model training process.

[0030] To solve this problem, the present invention introduces a new loss function measurement method - Normalized Wasserstein Distance (NWD). NWD is an indicator that measures the difference between two probability distributions. It can overcome the problem that traditional loss functions are sensitive to small target position deviations to a certain extent. Specifically, NWD measures the detection accuracy of the model by calculating the difference in feature distribution between the predicted bounding box and the true bounding box. Since NWD is highly sensitive to subtle changes in feature distribution, it can better capture the feature information of small targets, thereby improving the detection accuracy of the model.

[0031] In the HATSC-YOLOv10 network, NWD is used as part of the loss function to replace the traditional loss metric based on bounding box coordinates. Specifically, during the network training process, NWD is used to calculate the difference in feature distribution between the predicted bounding box and the true bounding box, and it is optimized as part of the loss function. By continuously reducing the value of NWD, the model can capture the feature information of small targets more accurately, thereby improving the detection accuracy and robustness of the model.

[0032] The overall architecture of the HATSC-YOLOv10 network is improved and optimized based on YOLOv10. Specifically, the network consists of three parts: feature extraction stage, feature fusion stage, and detection head stage.

[0033] The overall architecture of the HATSC-YOLOv10 network is a targeted improvement and optimization based on YOLOv10, aiming to improve the detection performance of small targets in remote sensing images. The network mainly consists of three core parts: feature extraction stage, feature fusion stage and detection head stage. The specific structure and function of each part will be described in detail below.

[0034] 1. Feature extraction stage The feature extraction stage is the basis of the HATSC-YOLOv10 network, which is responsible for extracting effective feature representations from the input remote sensing image. This stage uses multiple convolutional layers and pooling layers to gradually extract different scale features in the image through layer-by-layer convolution and pooling operations.

[0035] In the feature extraction stage, the hybrid attention transformer (HAT) module is specially introduced. The HAT module combines the advantages of channel attention and spatial attention, and can adaptively adjust the feature weights of different channels and spatial positions, thereby enhancing the model's ability to focus on local details in complex backgrounds. After each feature extraction stage, a HAT module is connected to perform dual adjustments of channel and spatial attention on the feature map, so that the network can more effectively extract and utilize small target features in remote sensing images.

[0036] In addition, in order to reduce the complexity and computational complexity of the model, the spatial and channel reconstruction convolution (SCconv) module is also used in the feature extraction stage. The SCconv module achieves efficient feature utilization and model compression by reconstructing and recombining the spatial and channel dimensions of the feature map. Some traditional convolutional layers are replaced by SCconv layers to reduce the parameters and computational complexity of the model while ensuring that the performance of the model is not greatly affected.

[0037] 2. Feature Fusion Stage The feature fusion stage is a key part of the HATSC-YOLOv10 network, responsible for fusing features of different scales and levels to extract richer and more effective feature representations. This stage uses a variety of feature fusion methods, including upsampling, downsampling, splicing, etc., to fuse feature maps of different scales to form a more robust feature representation.

[0038] In the feature fusion stage, special attention is paid to the fusion and utilization of multi-scale features. By introducing feature maps of different scales and fusing them, the network can better adapt to small target detection tasks of different scales. At the same time, advanced feature fusion methods such as feature pyramid network (FPN) are also used to further improve the effect and accuracy of feature fusion.

[0039] 3. Detection Head Stage The detection head stage is the output part of the HATSC-YOLOv10 network, which is responsible for object detection and classification based on the extracted feature representation. This stage uses multiple convolutional layers and fully connected layers to gradually transform the feature representation into the category and location information of the target through layer-by-layer convolution and fully connected operations.

[0040] In the detection head stage, the normalized Wasserstein distance (NWD) is specially introduced as part of the loss function. NWD is an indicator that measures the difference between two probability distributions, which can overcome the problem that the traditional loss function is sensitive to the position deviation of small targets to a certain extent. By using NWD as part of the loss function, the model can more accurately capture the feature information of small targets, thereby improving the detection accuracy and robustness of the model.

[0041] Specifically, during the network training process, NWD is used to calculate the difference in feature distribution between the predicted bounding box and the true bounding box, and is optimized as part of the loss function. By continuously reducing the value of NWD, the model can more accurately locate the position of small objects and give more accurate category prediction results.

[0042] In summary, the HATSC-YOLOv10 network has improved the performance of small target detection in remote sensing images by introducing innovative technologies such as the hybrid attention converter (HAT) module, the spatial and channel reconstruction convolution (SCconv) module, and the normalized Wasserstein distance (NWD) loss function. The network has the advantages of clear structure, efficient calculation, and high detection accuracy, and has broad application prospects in the field of remote sensing image target detection.

[0043] The above shows and describes the basic principles and main features of the present invention and the advantages of the present invention. It is obvious to those skilled in the art that the present invention is not limited to the details of the above exemplary embodiments, and the present invention can be implemented in other specific forms without departing from the spirit or basic features of the present invention. Therefore, no matter from which point of view, the embodiments should be regarded as exemplary and non-restrictive. The scope of the present invention is defined by the attached claims rather than the above description, and it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims are included in the present invention. Any figure mark in the claims should not be regarded as limiting the claims involved.

[0044] In addition, it should be understood that although the present specification is described according to implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation modes that can be understood by those skilled in the art.

Claims

1. An improved YOLOv10 network for small target detection in remote sensing images, characterized in that: include: A transformer (SC) module that combines a hybrid attention transformer and spatial and channel reconstruction convolutions for deep feature extraction and fusion of small objects in remote sensing images to enhance the feature extraction capability of local details under complex backgrounds; A YOLOv10 network structure, in which the converter SC module is integrated to form a converter SC-YOLOv10 network structure to improve the detection performance of small objects in remote sensing images; and a loss function in which NWD is introduced to measure the similarity between the bounding boxes of detected small objects to reduce the sensitivity of the loss function to small target position deviations.

2. The improved YOLOv10 network for small target detection in remote sensing images according to claim 1, characterized in that: The converter module is used to extract and fuse deep features of image details.

3. The improved YOLOv10 network for small target detection in remote sensing images according to claim 1, characterized in that: The channel reconstruction convolution is used to compress the detection network model and improve the performance of the detection network.

4. The improved YOLOv10 network for small target detection in remote sensing images according to claim 1, characterized in that: The NWD is used to measure the similarity between bounding boxes of small objects detected in the loss function, making the loss function insensitive to position deviations of small objects.

5. The improved YOLOv10 network for small target detection in remote sensing images according to claim 1, characterized in that: The converter SC-YOLOv10 network structure improves the performance of small target detection in remote sensing images by combining the converter SC module and the YOLOv10 network.

6. The method for improving the YOLOv10 network for small target detection in remote sensing images according to any one of claims 1 to 5, characterized in that: The following steps are involved: a. Use the converter SC module to perform deep feature extraction and fusion of remote sensing images; b. Input the extracted features into the YOLOv10 network integrated with the converter SC module for target detection; c. Use the NWD-introduced loss function to optimize the position of the detected targets to reduce the sensitivity to small target position deviations.

Images