Learning a Deep Convolutional Network for Speckle Noise Reduction in Underwater Sonar Images

Abstract

Underwater sonar imaging system has been widely utilized to detect and identify the submerged objects of interest. However, imaging quality often suffers from the undesirable signal-dependent speckle noise during signal acquisition and transmission. The speckle noise will restrict the practical applications, such as object detection, tracking and recognition, etc. To enhance the sonar imaging performance, we propose a deep learning approach to directly estimate the speckle noise in logarithmic domain based on the convolutional neural network. Once the speckle noise is obtained, the latent sharp image can then be easily calculated according to the image degradation model. The patch-based loss function, i.e., structural similarity metric, is adopted to preserve the important geometrical structures during speckle noise reduction. Experiments have been implemented on different noise levels to demonstrate the effectiveness of the proposed deep learning approach. Experimental results have illustrated that it outperforms several widely-used speckle noise reduction methods.