A Lightweight Serial CNN Model for Remote Sensing Ship Target Recognition on FPGA

Abstract

Ship target recognition with remote sensing is an important task for earth observation. The excessive recognition delay may render the results meaningless for time-sensitive targets such as ships. Therefore, rapid and timely analysis of image data on the space-borne platform is required. With the development of deep learning, Convolution Neural Network (CNN)-based methods have achieved high accuracy in the ship recognition task. The huge model size and computational complexity make the CNN-based methods difficult to be deployed on the space-borne platform with limited resources. Hence, model compression techniques are adopted in some on-board methods to obtain the real-time information of ships. However, these methods can only locate ships and cannot give specific category information. To address this problem, a lightweight ship recognition method based on the subgraph cascade classification is proposed. Firstly, the entire image is divided into small subgraphs. Secondly, a lightweight CNN is adopted to select ships from backgrounds. Then, the accurate coordinates of ships are obtained through position regression. At last, ship targets are classified as civil ships and other ships. The experiments show that the method has a promising application prospect in on-board intelligent processing.