On‐Orbit Fast Small Target Detection Method Based on the Improved YOLOv5 Model

Small target detection in remote sensing images is a challenging research field. Real‐time on‐orbit small target detection is of great significance for rapid discovery and tracking of remote sensing targets. In this paper, the feasibility of automotive‐grade system on chips (SoC) to achieve on‐orbit high‐performance computing is analyzed, and a high‐performance computing solution that meets the complex space application environment is proposed. At the same time, based on the improved YOLOv5s (You Only Look Once) model and the block parallel processing framework, a real‐time on‐orbit small target detection method for remote sensing images is proposed. Under the conditions of limited computing resources on orbit, this method ensures the effectiveness of target detection while improving the speed of target detection. Experimental results show that the small target detection time of this method is about 100 ms for a 5 × 3 k resolution image; meanwhile, the actual results for small object detection with the improved model are also better than the original YOLOv5 model.