In recent years, due to the importance of underwater image enhancement in underwater robot, underwater vehicle and ocean engineering, more and more extensive research has been done. It has evolved from implementing physics-based solutions to using very cutting edge cnn and GANs. However, these cutting-edge algorithms often come at the cost of high computing power and time, which reduces the efficiency and portability of underwater working equipment using these algorithms. At the same time, these models have harsh requirements on data sets, leading to high cost of training and unfriendly to many underwater operations. Therefore, this paper aims to propose a lightweight neural network structure, Shallow underwater neural network. These neural networks associate the original image directly with the output of each convolutional layer, preserving the original features while enhancing the image and avoiding gradient descent. The experimental results show that the model has a good effect on image enhancement, and the structure is lightweight.