A Ship Detection Model Based on YOLOX with Lightweight Adaptive Channel Feature Fusion and Sparse Data Augmentation

Abstract

Accurate and real-time detection of ships has become an essential part in maritime video surveillance and plays a vital role in national territorial water security. However, the existing ship detection models have poor recognition performance for some types of ships, such as small ships and ultra-long ships. Meanwhile, light and weather factors also affect the accuracy of the ship detection model, which blurs the boundaries between ships and the background In this paper, a new ship detection model ShipYOLOX based on YOLOX is designed to solve those problems in existing models. Specifically, the modelrs ability to discriminate ships with complex contours is enhanced by adding a feature fusion module called Lightweight Adaptive Channel Feature Fusion (LACFF). Additionally, a new data augmentation algorithm Sparse Target Mosaic is designed to replace the original Mosaic. The new model ShipYOLOX has been validated in experiments with results showing that it can improve the accuracy of ship detection, which achieves an excellent performance of 87.4% AP75 on Seaships.