An attention cut classification network for fine-grained ship classification in remote sensing images

Abstract

ABSTRACT Ship classification plays an important role in monitoring maritime regulations. Despite the practicality of fine-grained ship classification, ship classification is primarily done at a coarse level. To remedy this, we propose an attention cut classification network (ACCN). This network structure is designed with randomly cut images, attentionally cut images, and the raw images of both as inputs, which allows the network to focus on detailed features without excluding global features. Additionally, an attention classification module is designed to obtain more accurate local areas. The classification accuracy is improved using high-dimensional features. The performance of this model was evaluated using a fine-grained ship classification dataset, which is a public dataset containing 23 classes of ships. The ACCN performed better than previous classification models. When only raw images and category labels were used as inputs, the ACCN still achieved an overall accuracy rate of 93.67%. Furthermore, we applied this method to a remote sensing dataset and obtained good results.