Abstract: Aiming at the issue of low recognition accuracy for small targets in remote sensing images, which is caused by the presence of numerous small targets and complex backgrounds, an improved small target detection method for remote sensing images is introduced. Based on an enhanced YOLOv7 network model, this method incorporates a bi-level routing attention mechanism at the down-sampling stage to construct the feature extraction module for small targets MP-ATT (max pooling-attention). The model pays more attention to the characteristics of small targets and improves the accuracy of small target detection. To enhance the perception of small target details, two-dimensional convolutions in the backbone network are replaced with deformable convolution network v3 (DCNv3), and a new layer aggregation module ELAN-D is built. A novel small object detection layer is devised for the network to acquire more refined feature information, thereby enhancing the model’s robustness. Additionally, the loss function is optimized by replacing CIoU with MPDIoU (minimum point distance based IoU) in order to adapt to scale changes in remote sensing images and improve network robustness. Experimental results on the DOTA-v1.0 dataset demonstrate that the proposed method achieves significant improvement compared to the original model, the accuracy, recall and mAP(mean average precision) are improved by 0.4, 4.0 and 2.3 percentage points, which proves that the proposed method can effectively improve the detection effect of small targets in remote sensing images.

Key words: deep learning, target detection, remote sensing images, small target, YOLOv7

摘要： 针对遥感图像中小目标数量众多且背景复杂所导致的识别精度低的问题，提出了一种改进的遥感图像小目标检测方法。该方法基于改进的YOLOv7网络模型，将双级路由注意力机制加入至下采样阶段以构建针对小目标的特征提取模块MP-ATT（max pooling-attention），使得模型更加关注小目标的特征，提高小目标检测精度。为了加强对小目标的细节感知能力，使用DCNv3（deformable convolution network v3）替换骨干网络中的二维卷积，以此构建新的层聚合模块ELAN-D。为网络设计新的小目标检测层以获取更精细的特征信息，从而提升模型的鲁棒性。同时使用MPDIoU（minimum point distance based IoU）替换原模型中的CIoU来优化损失函数，以适应遥感图像的尺度变化。实验表明，所提出的方法在DOTA-v1.0数据集上取得了良好效果，准确率、召回率和平均准确率（mean average precision，mAP）相比原模型分别提升了0.4、4.0、2.3个百分点，证明了该方法能够有效提升遥感图像中小目标的检测效果。

关键词: 深度学习, 目标检测, 遥感图像, 小目标, YOLOv7