Multi-Scale Object Detection Algorithm Based on Decoupled Networks and Upsampling Aggregation

Multi-scale object detection requires addressing the challenge of differences between features at different scales. To improve detection precision across scales, a Scale Decoupled Networks (SDNets) and Upsampling Aggregation Module (UAM)-based algorithm is proposed. SDNets uses three parallel detection branches for small, medium, and large-scale objects to reduce conflicts between different scale features. Convolutions with different receptive fields extract features of different scales, enhancing the network's ability to express them. UAM adds a pathway for upsampling feature fusion during downsampling, improving the effect of feature fusion. Experimental results show that our algorithm achieves 83.2% detection precision on the VOC dataset, 2.0% higher than YOLOX and 0.3% higher than YOLOv7, and improves detection precision for small, medium, and large-scale objects by 3.7%, 2.7%, and 2.1%, respectively.