Multi class sea surface target dataset for autonomous navigation of ships

This article proposes a sea surface target dataset for autonomous navigation of ships, which includes 2290 basic images and 4 classifications. Selecting the optimal model for training and using the mesoscale experimental ship Chuangzhi for validation, the model accuracy reached 89.12%, the recall rate reached 80.28%, and the comprehensive index F1 score was 84%. The actual ship test has shown that the self built dataset has an average confidence increase of about 17% compared to the traditional COCO dataset, and is significantly better than the COCO dataset when facing multiple targets, small targets, and long-distance scenarios. This dataset can effectively improve the detection comprehensiveness, accuracy, and robustness of the model in actual sea surface scenarios, providing important support for the development of autonomous navigation technology for ships.