Adaptive spatial attention dual-branch fishing boat detection network

Abstract

Aiming at the harbor environment, the target detection accuracy of fishing vessels is low, and it is prone to the problems of fishing vessel misdetection and omission detection. In this paper, we propose a fishing vessel target detection algorithm called ASDNet based on YOLOX. Firstly, an Adaptive Spatial Attention Module (ASAM) was designed and used to improve the detection of fishing vessel targets; secondly, a two-branch backbone network was designed for multidimensional fishing vessel feature extraction. Meanwhile, a bilateral enhanced fusion strategy (BFFS) is designed to fuse the branch features to improve the characterization ability of the network; finally, the loss function is improved by introducing the Focal-CIOU loss bounding box loss function to reduce the effects of the detection position deviation of the fishing vessel target and the overlap of the vessel hull to improve the detection performance. The above methods are validated using the homemade fishing vessel dataset, and the results show that the precision rate (P) and recall rate (R) are greatly improved. The average precision rate (mAP@50-95) value reaches 80.25%, which is 2.39% higher than that of the 77.86% of the YOLOX. It significantly improves the precision of the detection, meets the requirements of the performance of the target detection of the fishing vessel, and has certain practical significance in engineering.