Abstract: In the target detection task of remote sensing images, in order to locate the target more accurately, the existing one-stage detection method based on the feature extraction of candidate frames is to fully preset multiple priori frames at each spatial position to cover the target to be detected. However, this will greatly increase the computational complexity of the one-stage detection method. This paper proposes a multi-directional remote sensing target detection method based on the feature correction of the candidate frame. In this method, only one candidate frame is preset at each position of the feature map. We replace the original box with the candidate box obtained after feature correction through regression learning, and then use the classification layer and regression layer of the one-stage detection method to identify and locate respectively. The method uses Mobilenetv2 as the basic structure of the detection network. The detection rate of aircraft on the DOTA dataset can reach 96.8%, the false alarm rate is 6.7%, the mAP value is 0.87, and it has complete real-time results, surpassing all remote sensing image detection methods based on candidate frame feature extraction such as SSD and YOLOv3. Because this method cleverly avoids the priori design of the aspect ratio and scale of the candidate frame, this method can be easily applied to other similar detection tasks, plug-and-play, it has strong task adaptability.

Key words: remote sensing image, target detection, feature correction