A Light-weight Ship Detection and Recognition Method Based on YOLOv4

Abstract

Ship detection and recognition based on deep learning often needs high standard hardware support while achieving high precision, which is difficult to adapt to offshore resource-limited platforms. Trying to solve this problem, this paper adopts the one-step target detection model YOLOv4 as the framework and applies a comprehensive network simplifying method. Firstly, this method applies different lightweight backbone networks in the framework to obtain the ideal Mobilenetv2-YOLOv4 network, and then conducts sparse training based on the scale factor of the batch normalization layer. Finally, it selects an appropriate threshold to prune unessential channel, which obtains a light-weight ship detection neural network for ship detection and recognition. The average accuracy of the network for detecting and identifying targets of 8 types of ships reaches 92.8% on average, the real-time detection speed is 37 frames per second, and the detection efficiency is 70% higher than that of the original network, which is capable of real-time detection under the condition of limited resources. The results also show that under simple tasks, appropriate methods can effectively compress the network parameters and computations while maintaining accuracy.