Application of Enhanced Feature Fusion Applied to YOLOv5 for Ship Detection
Shuaiyu Jin, Lei Sun · 2021
Ship detection is very important in improving navigation efficiency. In recent years, YOLO series object detection algorithm has achieved remarkable achievements. As the representative work of the YOLO series, version 5 (YOLOv5) is widely used in target recognition tasks for its high recognition accuracy and lightweight model. However, the sea scene background is complex and variable, which is greatly affected by extreme weather such as light, rain and fog. For this reason, an improved YOLOv5s algorithm with attention mechanism is proposed to enhance the feature fusion module. Firstly, the feature map extracted from different layers was aligned on the number of channels. Then, different weight coefficients were added to the corresponding feature map. Finally, the experimental results demonstrate that the improved algorithm obtain 3.1% higher mAP_0.5 than YOLOv5s on MS-COCO datasets. In the meanwhile, the performance on BOAT data is also competitive in terms of quality.