A Deep Ship Recognition Model Based on Multi-category Fusion of General Datasets

To meet the demand for accurate and rapid recognition of ship targets in existing ship image recognition methods, this study applies the DCNN model, known for its powerful feature extraction capabilities, to the task of ship image recognition. To enhance the model's performance, a novel dataset fusion training method is utilized. The DCNN model requires large-scale labeled training datasets; however, there is currently a lack of diverse and large-scale labeled ship image datasets in the public domain. Additionally, manually collecting and labeling large-scale ship recognition data is labor-intensive. Therefore, this paper proposes a new approach by creating a ship target recognition dataset that fuses existing general datasets for training the ship recognition model. Furthermore, a model training method that integrates multi-category image recognition is proposed. In this method, a focal loss-based optimization training approach is specifically employed to handle imbalanced data classification, thereby improving the recognition performance of ship images within multi-category datasets. Performance tests on the trained model show that the mean Average Precision (mAP) of the model trained with multi-category fusion is relatively improved by $9.45 \%$ compared to the $\mathbf{3 8 . 1 9 \%}$ mAP of the model trained without focal loss on a single category. Experimental results demonstrate that the proposed multi-category image training method can effectively enhance the performance of ship image recognition tasks in training using existing general image recognition datasets.