Lunar Surface Target Detection Based on Improved YOLOv8

In lunar exploration and research, the accurate identification and localization of lunar surface features is of great scientific significance and application value. However, target detection in lunar surface images faces a series of difficulties. Firstly, the lunar surface rock and soil is relatively monotonous and lacks texture information. Secondly,amount of image data on the lunar surface is less compared to most of the image data and there are overlapping areas between the images. Therefore, manually labeling lunar surface features is susceptible to significant errors. Thirdly, the size of the target in the lunar surface images is small and the resolution of the panorama spliced images of the lunar surface is high, leading to small target detection on high-resolution images a major technical difficulty. Aiming at the difficulty of data annotation of lunar surface images, this paper proposes a data annotation method based on panoramic spliced images. Then an improved YOLOv8 target detection algorithm is proposed by introducing an EMA(Expectation-Maximization-Attention) attention mechanisms, an adaptive feature fusion module and a sliding window based slice detection strategy, which solves the difficulty of detecting small targets on high-resolution images and avoids the problems of false detection and missed detection.