A Convolutional Neural Network Based on Double-tower Structure for Underwater Terrain Classification

Terrain classification plays a critical role in all robot systems especially in unknown environments. In recent years, researchers have proposed various algorithms to improve the efficiency and accuracy of terrain classification. Nevertheless, these methods still have some deficiencies in classification efficiency. In this paper, a double-tower convolutional neural network has been designed to implement end-to-end underwater terrain classification. The matched sonar image and visual image constitute an image pair, which is obtained at the same time by the sonar sensor and the visual sensor of the robot or underwater vehicle. The corresponding image pairs are set to be the input of the convolutional neural network, and the output of the network is the classification of the terrain. Then, terrain features from sonar and visual images are simultaneously applied to achieve terrain classification. Therefore, an end-to-end convolutional neural network with a classification function has been established in this paper.