Ship Detection with Optical Image Based on CA-YOLO v3 Network
Danmeng Li, Zhuo Zhang, Zhengwei Fang, Fuxiao Cao · 2023
This article proposes an improved ship target detection network, CA-YOLO v3 Network, which is based on the YOLO v3 algorithm. The aim is to address the challenge of identifying ship categories and positions in complex sea environments, especially when multiple ships occlude each other. Firstly, the ConvNeXt module is used to replace the Darknet53 network in the backbone network, which solves the problem of insufficient image extraction for multiple ship targets, avoids the loss of feature information and reduces the floating-point operation of the network. Secondly, to fully utilize feature information of different scales, the ASFF module is introduced between the output feature layer and detector to achieve feature scale fusion and size unification, which improves the model detection accuracy while ensuring model lightweight. During training, data enhancement is performed on the dataset at the input side, and the GIOU loss function and 9 new anchors matching the ship type are used for training, enabling faster and more accurate determination of the anchor box size and position. Experimental results show that compared to the original YOLO v3 network, the mAP of the CA-YOLO v3 network increases from 74.76% to 95.12%, achieving a significant improvement of 27.2%.