Ship fine-grained classification network based on multi-scale feature fusion

Ships are essential maritime transportation carriers and military targets. Utilizing remote sensing technology to accurately and automatically identify and classify them has broad application prospects and significant practical implications. The challenge in fine-grained ship image classification lies in the subtle differences between various types of ships and significant variations within the same type. This challenge is further compounded by the scarcity of available ship datasets. In order to tackle these issues, we have put forward a ship image classification model that utilizes multi-scale feature fusion for more precise analysis. This research used a multi-scale feature enhancement module to extract feature representations from multi-scale ship images and enhance the intra-class correlation in ships. Subsequently, the multi-scale feature fusion module combined the features to get global correlations of ship images as well as local features. Additionally, this research established a ship dataset comprising six common categories. Ultimately, we assessed the suggested approach on both the FGSCR-42 dataset and our own ship dataset. The results indicate that when compared to ResNet-50, this method significantly improved the fine-grained classification accuracy of ships. The accuracy, precision, recall rate, and the F1Score index reached 98.96%, 97.58%, 97.54%, and 97.56%, respectively.