NAS-YOLOX: ship detection based on improved YOLOX for SAR imagery
Hao Wang, Dezhi Han, Zhongdai Wu, Junxiang Wang, Yuan Fan, Yachao Zhou · 2023
Synthetic aperture radar (SAR) satellites can provide microwave remote sensing images that are not limited by weather and light, so they are widely used in the field of ocean monitoring. The current SAR ship detection method based on deep learning (DL) is difficult to more effectively fuse complex features, which leads to low detection accuracy of target ships and even missed or false detections. In order to solve this problem, this paper proposes an improved YOLOX-based SAR image ship detection method, called NAS-YOLOX. Based on the YOLOX algorithm, path aggregation feature pyramid network (PAFPN) is replaced by a neural architecture search - feature fusion network (NAS-FPN) to enhance the cross-scale fusion ability of the proposed model. And a dilated convolution feature enhancement module (DFEM) is also designed and embedded into the backbone network to boost the network receptive field and the ability to extract target information. Furthermore, a multi-scale channel-spatial attention (MCSA) is proposed to improve the attention to key areas of the ship. The experimental results on the HRSID public data set show that the AP0.5of NAS-YOLOX is 6.3% higher than that of the YOLOX model. Compared with other ten mainstream target detection algorithms, NAS-YOLOX has also achieved excellent detection result.