Ship Detection with Lightweight Network Based on YOLOV3

This paper proposes a lightweight network for ship detection based on YOLOV3. This method reduces the parameters and model size of YOLOV3 while retaining accuracy from three aspects. First, an efficient way is described of replacing the Darknet-53 of YOLOV3 with a reduced network. Second, an optimization form is demonstrated to improve the basic unit. Last, compression and acceleration strategies are implemented in the method. The performance of our method and YOLOV3 is measured on a self-built dataset for ship detection. The method we proposed is 29x less parameters, 85x less size, and 3 x less inference time than YOLOV3. In the experiment, the method is deployed on a Qihang Unmanned Surface Vehicle (USV) and realizes real-time detection of ships.