Muyuan Fan,Ziheng Xia,Jinhua Wei,Yingxia Wang,Bingxiu Yang,Wenjun Huo
出处
期刊:日期:2025-10-24卷期号:: 1107-1111
标识
DOI:10.1109/icsps66615.2025.11347771
摘要
The YOLO series exhibits unique advantages in remote sensing image target recognition tasks due to its end-to-end efficient inference mode and real-time detection capabilities. However, it still struggles to fully address the inherent challenges of remote sensing target detection. For instance, the traditional Spatial Pyramid Pooling (SPP) module lacks explicit feature enhancement, resulting in the loss of small target information. Additionally, attention mechanisms inadequately model the long-range dependencies between targets and complex backgrounds in remote sensing scenes.To tackle these issues, this paper proposes a You Only Look Once network based on Channel and Spatial Multi-scale Fusion (CSMF-YOLO), which integrates the SPPF_Fusion and Spatial - Channel Collaborative Attention(SCCA) modules. Experimental results on the DOTA remote sensing dataset demonstrate that this method improves the mean average precision (mAP) by $4.8 \%-7.8 \%$ compared to the original YOLOv8, with particularly significant performance improvements in small target detection and complex background scenarios. The innovations of this paper lie in constructing a hierarchical processing framework of “feature enhancement -multi-scale fusion - attention optimization.” It is the first to combine multi-scale pooling with explicit feature enhancement to handle the scale variations of remote sensing targets. By using adaptive branch attention, it resolves the problem of weight allocation in small target feature fusion. Moreover, it enhances the global contextual understanding ability through the collaborative modeling of spatial and channel attention, providing a superior technical solution for small target detection in remote sensing images.