GLFE-YOLOX: Global and Local Feature Enhanced YOLOX for Remote Sensing Images

Object detection for remote sensing images (RSIs) is an important research topic in remote sensing data analysis. Many efforts have been devoted to remote sensing object detection (RSOD) tasks, most of which try to use attention mechanisms to improve the performance of detectors. However, the difference information between the global features and local features of feature maps is ignored. In this paper, we design a novel global and local enhanced attention mechanism (GLE-AM) to capture this difference information. Then, we propose a global and local feature enhanced network (GLE-Net) to fully utilize the features extracted by the GLE-AM. Furthermore, in order to make the path aggregation feature pyramid network (PAFPN) more suitable for extracting fused features and small object detection tasks, we improve the cross stage partial layer (CSP-Layer) and the spatial pyramid pooling (SPP), respectively. Experiments conducted on two publicly available remote sensing datasets demonstrate the effectiveness of our proposed methods. On the DIOR dataset, GLFE-YOLOX improved on the mAP metric by 3.19% compared to YOLOX-m baseline, and on the NWPU VHR-10 dataset GLFE-YOLOX reaches 90.93% on mAP, which is 3.18% higher than YOLOX-l, comparing with the comparison algorithms, our proposed GLFE-YOLOX achieves the best results in mAP metric for both datasets.