YOLO-Parallel: Positive Gradient Modeling for Long-Tail Remote Sensing Object Detection

The long-tail distribution problem is widely prevalent in remote sensing images (RSIs), posing significant challenges to object detection tasks. Most existing methods for long-tail detection are designed for two-stage models. Such approaches of suppressing negative gradients tend to increase false alarms in one-stage detectors, resulting in a decline in overall performance and an increase in postprocessing time. This letter presents a novel long-tail loss with broad applicability in diverse you only look once (YOLO) networks. We present a novel positive gradient loss (PGLoss) that effectively enhances the accuracy of tail categories while preserving the accuracy of head categories. Furthermore, to address the performance degradation caused by the pseudo-residual structure, we create parallel block with efficient computation and superior feature extraction abilities. We designed and trained the network named you only look once (YOLO)-Parallel to verify the effectiveness of PGLoss and parallel block. Extensive experiments were conducted on two large-scale optical remote sensing datasets, DIOR and DOTA, which are severely affected by the long-tail problem. The results powerfully demonstrate the superiority of our algorithm. YOLO-Parallel, with only 33.3% of the parameters of YOLOX, achieved a comparable detection performance of 96.9% on DIOR. On the DOTA dataset, PGLoss achieved mean average precision (mAP) improvements of around 1.5% for YOLO-Parallel, YOLOv5n, and YOLOv7-tiny without increasing NMS processing time.