Scholarlink

<p>In the field of remote sensing image target detection, some images are difficult to identify due to small targets and inconspicuous features. Traditional models have low detection accuracy, and greatly improving detection accuracy can lead to excessive computation and slow processing speed, resulting in a decrease in real-time performance. Therefore, a method for detecting objects in remote sensing images based on GCC-YOLOv5 is proposed. Firstly, in YOLOv5s, the lightweight GSConv structure is used to replace convolutional layers in the backbone and head to reduce redundant parameters; Secondly, the Convolutional Block Attention Module is added to the model to improve the detection effect of small targets; Then, replace the upsampling module in Head with Content-Aware ReAssembly of Features upsampling operator to improve partial detection accuracy while maintaining lightweight network; Finally, the WIoU loss function is introduced to further improve detection accuracy. The experimental results show that the improved GCC-YOLOv5 model has a&#xa0;better small target detection effect than the original yolov5 model. On the DIOR dataset, P, R and mAP@0.5 reached 89.1%, 79.7% and 84.9%, respectively, which were 3.6%, 1.0% and 2.0% higher than the original network. The FPS reached 113.0, an increase of 80.0% compared with the original network. The Params and FLOPs are 12.5M and 13.7G, respectively, which are 14.4% and 13.8% lower than those of YOLOv5s. The proposed approach demonstrates enhanced detection performance on both the NWPU VHR-10 and SAR-Airport-1.0 datasets, while simultaneously minimizing model parameters and computational overhead. The mAP@0.5 achieved on the two datasets is 92.2% and 99.3%, respectively.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

GCC-YOLOv5: An Algorithm for Object Detection in Remote Sensing Images

  • 期刊论文

  • 发表日期：

    2025年5月14日

查看全文
  • Chengjie Zhu,
  • Hao Jiang

摘要

In the field of remote sensing image target detection, some images are difficult to identify due to small targets and inconspicuous features. Traditional models have low detection accuracy, and greatly improving detection accuracy can lead to excessive computation and slow processing speed, resulting in a decrease in real-time performance. Therefore, a method for detecting objects in remote sensing images based on GCC-YOLOv5 is proposed. Firstly, in YOLOv5s, the lightweight GSConv structure is used to replace convolutional layers in the backbone and head to reduce redundant parameters; Secondly, the Convolutional Block Attention Module is added to the model to improve the detection effect of small targets; Then, replace the upsampling module in Head with Content-Aware ReAssembly of Features upsampling operator to improve partial detection accuracy while maintaining lightweight network; Finally, the WIoU loss function is introduced to further improve detection accuracy. The experimental results show that the improved GCC-YOLOv5 model has a better small target detection effect than the original yolov5 model. On the DIOR dataset, P, R and mAP@0.5 reached 89.1%, 79.7% and 84.9%, respectively, which were 3.6%, 1.0% and 2.0% higher than the original network. The FPS reached 113.0, an increase of 80.0% compared with the original network. The Params and FLOPs are 12.5M and 13.7G, respectively, which are 14.4% and 13.8% lower than those of YOLOv5s. The proposed approach demonstrates enhanced detection performance on both the NWPU VHR-10 and SAR-Airport-1.0 datasets, while simultaneously minimizing model parameters and computational overhead. The mAP@0.5 achieved on the two datasets is 92.2% and 99.3%, respectively.