【摘要】 针对遥感图像中小目标被漏检、错检的情况，提出基于改进YOLOv7的小目标遥感图像识别算法。通过构建坐标卷积堆叠分支CCSB(Coordinate Convolutional Stacking Branch)替换原ELAN,提升网络对密集目标的检测能力；基于浅层特征提出联合特征提取模块JFEM(Joint feature extraction module),提取多尺度信息；提出深层路由注意力模块DRAM(Deep routing attention module),使模型聚焦于图像中的关键信息；基于信息融合提出特征融合策略FFS(Feature fusion strategy),去除特征金字塔内部冲突信息；提出混合损失函数MLF(Mixed loss function),提升对目标的定位能力。结果表明：在DIOR遥感数据集上mAP达到92.32%,较原始YOLOv7提高了3.63%;在RSOD数据集上mAP达到97.80%,较原始YOLOv7提高了3.50%,证明了所改进方法在遥感图像上的有效性。更多还原

【Abstract】 In view of the situation where small targets in remote sensing images are missed or wrongly detected, a small target remote sensing image recognition algorithm based on improved YOLOv7 is proposed. Construct a coordinate convolution stacking branch CCSB(Coordinate Convolutional Stacking Branch) to replace the original ELAN to improve the network’ s detection ability of dense targets; propose a joint feature extraction module JFEM(Joint feature extraction module) based on shallow features to extract multi-scale information; propose the DRAM(Deep routing attention module) to make the model focus on the key information in the image; a feature fusion strategy FFS(Feature fusion strategy) is proposed based on information fusion to remove conflict information within the feature pyramid; a mixed loss function MLF(Mixed loss function) is proposed to improve the positioning ability of the target. The results show that the mAP reaches 92.32% on the DIOR remote sensing data set, which is 3.63% higher than the original YOLOv7. The mAP reaches 97.80% on the RSOD data set, which is 3.50% higher than the original YOLOv7,which proves the effectiveness of the improved method on remote sensing images.更多还原