Ship Target Detection Algorithm Based on Improved YOLOv3

Abstract

In view of the increasingly complex maritime traffic, the low level of intelligence in safety supervision, the low recognition rate and poor real-time performance of current ship detection algorithms. In this paper, we propose an improved YOLOv3 algorithm, which can better detect ship targets on the sea. First, by replacing IOU in YOLOv3 with Complete_IOU, keeps the prediction box close to the ground truth center point and increases the recall rate when non-maximum suppression (NMS) is used; The prior anchor box size is redesigned using K-means++ clustering algorithm and evenly distributed to the corresponding prediction scale; In the loss function design, the previously improved Complete_IOU loss is used to replace the mean square error (MSE) loss, which solves the inconsistency of optimization and scale sensitivity. And improves the confidence loss function to solve the problem of positive and negative sample imbalance. Comparing this algorithm with other target detection algorithms on ship dataset, the experimental results show that the mean accuracy (mAP) of this algorithm on ship images is 89.84%, which is 4.57%, 7.23% and 3.75% higher than Faster R-CNN, SSD and traditional YOLOv3 algorithms respectively, and the frame rate is 30fps, which enables real-time target detection. Experiments show that this method can detect ship target accurately in real time which meets the needs of practical application.