Research on ship detection technology based on improved YOLOv5

Abstract

Ocean scene perception is the premise for unmanned ships to effectively complete all kinds of established tasks, and ship detection is the basic task of perception. Improving the accuracy of marine ship detection algorithms is of great importance to improve the working ability of unmanned ships. Due to the complexity of the marine environment, the data set that can be used to detect ships on the sea is small. On behalf of solving the mentioned problems, this paper suggests an algorithm based on YOLOv5 according to the characteristics of visible image ship detection in the unmanned ship perception system, optimizes the input end, loss function and detection box of the depth learning network model, and uses the migration learning strategy to train the network model. The experimental results manifest that the average precision (AP) of the algorithm for ship detection in the sea surface visible image reaches 98.6%, 1.69 percentage points higher than YOLOv5, and the average detection time per picture is about 45ms, which can meet the demands of ship detection in different situations.