Image Preprocessing and YOLO Architectures for Enhanced Small and Slow-Moving Object Detection
Diana Velychko, Saurav Singh, Panos P. Markopoulos, Eli S. Saber, Jamison R. Heard · 2024
This paper enhances the detection of small and slow-moving objects in satellite video imagery by integrating classical signal processing techniques, such as Accumulative Multiframe Differencing (AMFD) and Low-Rank Matrix Completion (LRMC), with deep learning models. We conduct experiments on the Video Satellite Objects (VISO) dataset using YOLOv5, YOLOv8, and YOLOv10 models. Notably, AMFD outperformed LRMC and a pre-trained YOLOv5, achieving a precision of 0.540, recall of 0.210, and $F 1$-score of 0.300. Furthermore, a YOLOv10 model trained from scratch on VISO for 250 epochs demonstrated superior performance, with a precision of 0.766, recall of 0.334, and $F 1$-score of 0.465. Low-resolution images ($220 \times 286$ pixels) achieved the highest precision (0.990) and $\mathbf{F 1}$-score (0.427). This study underscores the challenges in satellite imagery object detection, particularly regarding domain adaptation and resolution impacts, and paves the way for more effective object tracking.