LCE-Net: Local-Aware and Context Enhancement based YOLOv5 for object detection in remote sensing images

Remote sensing image target detection has been a research hotspot in the field of remote sensing. Aiming at the problems of complex background of remote sensing images, few pixels and large scale variability of remote sensing targets, a Local-Aware and Context Enhancement network(LCE-Net) is proposed with YOLOv5m as the baseline model. Firstly, the context enhancement module is designed in the network extraction layer to increase the perceptual field to fully extract feature information. Secondly, a cascade Swin Transformer block is added at the detection to capture feature information of object in similar environments. Thirdly, Alpha-CIoU to improve the localization accuracy. We validate the remote sensing image target detection algorithm on the DOTA dataset and the Plane dataset. The experimental results show that our algorithm increases the overall mAP from 69.4% to 73% compared to the YOLOv5m algorithm, which improves the remote sensing image target detection performance.