ShipYOLO: An Enhanced Model for Ship Detection

Abstract

The application of ship detection for assistant intelligent ship navigation has stringent requirements for the model’s detection speed and accuracy. In response to this problem, this study uses an improved YOLO-V4 detection model (ShipYOLO) to detect ships. Compared to YOLO-V4, the model has three main improvements. Firstly, the backbone network (CSPDarknet) of YOLO-V4 is optimized. In the training process, the 3 × 3 convolution, 1 × 1 convolution, and identity parallel mode are used to replace the original feature extraction component (ResUnit) and more features are extracted. In the inference process, the branch parameters are combined to form a new backbone network named RCSPDarknet, which improves the inference speed of the model while improving the accuracy. Secondly, in order to solve the problem of missed detection of the small-scale ships, we designed a new amplified receptive field module named DSPP with dilated convolution and Max-Pooling, which improves the model’s acquisition of small-scale ship spatial information and robustness of ship target space displacement. Finally, we use the attention mechanism and Resnet’s shortcut idea to improve the feature pyramid structure (PAFPN) of YOLO-V4 and get a new feature pyramid structure named AtFPN. The structure effectively improves the model’s feature extraction effect for ships of different scales and reduces the number of model parameters, further improving the model’s inference speed and detection accuracy. In addition, we have created a ship dataset with a total of 2238 images, which is a single-category dataset. The experimental results show that ShipYOLO has the advantage of faster speed and higher accuracy even in different input sizes. Considering the input size of 320 × 320 on the PC equipped with NVIDIA 1080Ti GPU, the FPS and mAP@5 : 5:95 (mAP90) of ShipYOLO are increased by 23.7% and 13.6% (10.6%), respectively, with an input size of 320 × 320, ShipYOLO, compared to YOLO-V4.