Fast ship detection based on lightweight YOLOv5 network

Aiming at a series of problems such as detection accuracy, calculation blocking, display delay, and so on in the ship surveillance video, an improved YOLOv5 algorithm is proposed this paper. First, to improve performance, it optimize anchor box network according target characteristics. Then, t-SNE used reduce visualize data set label information perform weighted analysis processed features for low-dimensional data. The mapped kernel k-means clustering adaptively selects more appropriate considers performance large small targets. Secondly, problem computational blocking BN scaling factor ? compress network, that model can be reduced without reducing performance. optimized framework trained self-integrated set. accuracy increased by 2.34%, speed reaches 98 fps 20 server environment low computing power version (Jetson nano), respectively.