Title: Underwater objects classification method in high-resolution sonar images using deep neural network

Author(s): ZHU Keqing; TIAN Jie; HUANG Haining;

Affiliation(s): Institute of Acoustics, Chinese Academy of Sciences; Key Laboratory of Science and Technology on Advanced Underwater Acoustic Signal Processing, Chinese Academy of Sciences; University of Chinese Academy of Sciences

Abstract: To solve the problem of underwater proud object classification using high-resolution sonar image under small sample situation, a classification method using deep neural network is proposed. Firstly, statistical characteristics of acoustic shadow regions are modeled using Gaussian mixture model and acoustic shadow is extracted. Trial and simulated dataset are constructed on this basis. Then, simulated dataset is input into convolutional neural network for training, and the feature extraction part is retained, which is used to extract feature of trial dataset. The classification part is reconstructed and trained by feature vectors of trial dataset. The experimental results show that the average classification accuracy of the proposed method is 88.24%,which is 8.67%,20.47%,19.78%,11.59%,9.01%,11.58% higher than that of other six methods respectively. It verifies that the proposed method achieves better performance on underwater proud object classification problem. The learning curve converges to 96.25%, which is 5.14% higher than validation curve, indicating that the over-fitting problem is alleviated to some extent. Improved convolutional neural network is applied in a fusion classifier, which also combines output of logistic classifier, support vector machine, and finally obtains a fusion result. The classification accuracy is up to 93.33%,indicating that fusion classifier improves robustness and classification performance of algorithm further. The proposed method combines deep learning and transfer learning, which not only utilizes powerful image classification ability of convolutional neural network, but also avoids serious over-fitting problem caused by limited dataset.