AM-ResNet: An Attention-Based Multi-Label Classification Network

Abstract

Exploring sea ice types in complex marine environments is essential to ensure the navigation safety of maritime activities like ships, offshore RIGS, and fishing in ice areas in winter. In this study, we propose a multi-label sea ice classification model based on the attention mechanism, AM-ResNet. First, we develop a multi-label classification model of neural networks based on features of sea ice structure and developmental processes. We then introduce a channel attention mechanism, the SE module, to improve the identification accuracy of the model. Finally, we embed the SE module into the classification model and construct a multi-label classification network based on attention mechanism and used it for sea ice multi-standard type classification. Based on the experimental results of the test set, the average classification accuracy reached 0.952, including 0.95 based on the texture features and 0.954 based on the development process features. This method achieves multi-label sea ice classification with high accuracy.