Enhancement for underwater image has been gradually valued with the development of ocean engineering as well as remote operated vehicles. Multifarious methods are applied in underwater image enhancing recently. Particularly, varieties of convolutional neural networks (CNN) have been applied for this field. However, during distrinct conditions lightness and medium quality, underwater images are complex and variable, which makes it much more challenging for CNN models to enhance images in different environments. In this paper, improvements are made to the architecture in Water-Net, in which we increased the enhancement units (E-Units) in the backbone of the network and improved the output of the confidence map. Experimental results on Underwater Image Enhancement Benchmark (UIEB) indicate that our network achieved better capability than other algorithms.