A super-resolution reconstruction method of underwater target detection image by side scan sonar

Abstract

However, the scope and distance of optical imaging were limited, especially in the case of muddy water, the propagation of optical information was seriously interfered, and imaging became more difficult. Due to the complex and changeable underwater environment and the nature of acoustic imaging, sonar image has noise, low resolution and fuzzy details, which has a great impact on the recognition and interpretation of sonar image. On the basis of the original SRGAN network, this paper improves and optimates its network structure and loss function. Replace the ordinary convolution layer with the void convolution layer in the residual block structure of the generated network, delete the batch normalization layer (BN layer), reduce the resource consumption and expand the receiver field, so as to improve the training efficiency of the network; A gradient penalty term is added to the improved discriminant network loss function to accelerate the convergence of the network and improve the stability of training. Four classical image super resolution algorithms are compared with the improved SRGAN algorithm under the verification of sonar dataset. The experimental results show that the improved SRGAN network is superior to the traditional super resolution method in the reconstruction of sonar image in terms of rich texture and details, and improves the quality of sonar image super resolution reconstruction.