Semantic Segmentation of Sea Ice Based on U-net Network Modification

Sea ice detection is essential to ensure the safe navigation of ships in mid- and high-latitude ice areas. In the face of complex sea ice information, how to use the sea ice images by shipboard cameras to comprehensively, accurately and efficiently identify four types of sea ice information (Ice skin, Nile ice, Grey ice and White ice)and two kinds of sea ice background information (sea water and sky), it remains a major challenge. This study proposes an automatic semantic segmentation method for sea ice images, which first uses Rsenet50 as well as Vgg-16 network to pre-train the model and improve the network training efficiency. Then modifications to U-Net network, improvement of the coding phase of the U-Net by introducing Vgg-16 and the residual structure, construction of the new network RU-Net and VU-Net. Compared with traditional classification methods, the experimental results show that the network can accurately identify all sea ice information in sea ice images. In particular, multi-scale sea ice types can be identified in real time, greatly improving the efficiency and accuracy of the identification of sea ice types. The MIoU values were 0.73 and 0.87 and the MPA values were 0.87 and 0.94 respectively.