Underwater Image Enhancement Based on Improved Water-net Model

Complex underwater environments and scarce real-world training datasets pose substantial obstacles to underwater image enhancement models based on Convolutional Neural Networks(CNN). In this paper, we first optimize the main network framework for the Water-net baseline model by selecting two convolutional blocks that differ only in the number of channels to be learned. Then in the convolution block connection process, we concatenate the original input with the convolutional output to enhance feature integration. For fusion, three different and more advantageous preprocessing methods are selected and increases the AquaBlend ConvModule (ABCM) enhancement module to enhance output quality while preventing overfitting. The improved model shows intentional performance on the full reference metrics, which is significantly improved compared to the baseline model, and the effectiveness of our method is verified in experiments on diverse underwater datasets.