Abstract: Target detection related technology has been widely used in space target surveillance, satellite automatic orbit finding and other fields. It is also one of the most important and challenging research branches in the field of computer vision and has gradually become a hot research area in the military field at home and abroad. In the modern air space confrontation, the acquisition of near earth vehicle targets through satellite remote sensing images can quickly judge the effective strength of the enemy forces, which enables our troops to occupy a strategic advantage. Aiming at the problems such as too long satellite observation distance and complex background of remote sensing image, the small sample target detection algorithm is studied based on one stage light weighted network YOLOv8. The research work of this paper mainly includes three aspects. Firstly, the generalization performance of the model is improved by image enhancement methods such as image flipping, Mosaic data enhancement and mixup data enhancement. Secondly, the average accuracy of the model is improved by adjusting the optimization function, reducing the class loss gain and reducing the mask ratio. Thirdly, the computational efficiency of the model is improved via the preset parameter and loading the model derived from the original optimization function. The method proposed in this paper is verified on the public aircraft data set, and the verification indexes include precise recall (PR), average accuracy (mAP) and the number of frames per second (FPS). The results show that the improved network model proposed in this paper can meet the needs of fast target detection in satellite remote sensing images.

Key words: aircraft image, satellite data, convolutional neural network, YOLOv8, fast object detection