Objective The research on the detection and recognition technology of marine ships has important economic and strategic significance for conservation of maritime rights and marine safety. This research is bounded and driven by several complex marine scenarios, especially severe weather conditions such as dense fog due to abundant vaporization of sea water and appearance of large areas of scale light and irregular water reflections with bright illuminations on the sea surface. Furthermore, ship targets become very small for large-scale sea and sky scenes, occlusions may also be seen near ports due to dense distribution of ship targets. All these problems make traditional background modeling techniques in the space-time domain unsuitable for marine target detection and recognition. In this research, the problem of ship target detection and recognition in sea and sky scenes is studied.

Method To address the problem of low accuracy and high misdetection rates caused by small targets, densely distributed targets and other occlusions; this research work proposes a novel ship target detection and recognition method based on the improved YOLOv3 algorithm. The feature extraction network of the YOLOv3 algorithm has been improved to enhance the feature extraction ability. Subsequently, the fusion method of Res2Net and SE-Net is adopted to enhance the robustness and generalization ability of the algorithm model and the output prediction scale of the model is expanded to effectively improve the accuracy of detection and recognition of small ship targets. By increasing the weight for small ship targets in the loss function, the misdetection rate of small ship targets is further reduced. Furthermore, through large number of comparative experiments, the performance of the ship targets detection and recognition method based on the improved YOLOv3 algorithm is tested on small ship targets, dense ship targets and ship targets with severe occlusions.

Results Experimental results show that the improved YOLOv3 algorithm can effectively achieve accurate detection and recognition of ship targets in the above scenarios, and the mAP based on improved YOLOv3 algorithm proposed in this paper increased 9.0% than YOLOv3 algorithm.

Conclusion The improved YOLOv3 algorithm proposed in this paper can relatively well solve the problems of high detection omission rate caused by small ship targets, dense distribution, and occlusion in large–scale maritime scenarios.