Improved Shallow-UWnet for Underwater Image Enhancement

Underwater image enhancement technology is significant to the development of underwater monitoring and autonomous underwater robots, and has always been one of the hotspots of underwater image enhancement research. At present, deep learning methods have been utilized to process image, in which heavy models such as GANs and deep CNNs will consume a lot of memory and numerical cost in the calculation task due to the large volume of them, thus lowering the efficiency in underwater tasks. In this paper, on the basis of Shallow-UWnet, an improved model is proposed which contains a series of convolutional blocks, batch normalization layers and LeakyReLU activation function. Further, a combination of mean squared error loss, perceptual loss and structural similarity loss is used. Based on the trained model, we compare the performance with three other advanced methods on three evaluation metrices and demonstrate our model's superior performance and generalization ability. Moreover, we test the proposed model's effect on two engineering cases and highlight its practical meaning.