Remote Sensing Image Target Detection Model Based on Improved YOLOX

In order to solve the problems such as complex and diverse backgrounds, dense object distribution and significant differences in target scale in high-resolution remote sensing image target detection, such as false detection and missing detection, this paper proposes a remote sensing image target detection network based on YOLOX, which realizes the high precision identification of small and medium targets in remote sensing image. Firstly, the SimAM attention module is introduced into three feature layers of YOLOX backbone network output to improve the attention of essential features and restrain the interference of complex backgrounds. Then, the loss function is optimized, and the binary cross entropy loss is replaced by Focal Loss, which increases the weight of hard-to-classify samples and the detection ability of dense targets. Finally, a comparative test is carried out on the publicly available remote sensing image target detection dataset NWPU-VHR-10, and good results are achieved. Compared with the benchmark network (YOLOX), the proposed algorithm improves the map index by 1.1%. The experimental results show that the proposed method can significantly improve the remote sensing small target detection performance and is superior to other advanced algorithms.