Optimizing the loss function for bounding box regression through scale smoothing

Abstract

Deep learning technology is widely used in target detection in machine vision. However, existing regression loss functions used for training networks suffer from slow convergence and imprecise localization, hindering the realization of fast and accurate visual detection. To address this, the study proposes the smoothing adaptive intersection over union loss (SAIoU Loss), which adapts bounding box regression through scale smoothing. By analyzing the bounding box regression process, SAIoU Loss incorporates a center-of-mass distance penalty term to enhance prediction speed during box distance regression in the pre-training phase. Additionally, it integrates a corner point distance penalty term with adaptive weights to refine the similarity of predicted box shapes throughout regression. The experimental results demonstrate that SAIoU Loss achieves a 39.6 mAP in target detection model training on PASCAL VOC, marking a 3.39% improvement. It also records the highest result of 26.7% in medium-sized target detection, which represents a 9.43% improvement over IoU. In the VisDrone 2019 dataset, SAIoU Loss reaches a detection accuracy of 14.8 mAP, improving by 1.3 mAP compared to the Baseline. The SAIoU loss proposed in this study realizes efficient and highly accurate target detection.