To improve the qualities of underwater laser images, the generator network of the generative adversarial network is changed to be a deep convolution neural network which contains the jumping structure and the dilated convolution. The network is used to learn the end-to-end parameters which map the unrepaired images to the target images from the self-built data set and repair underwater laser images with strong backscatter light. Compared with the experiment results of the joint processing of the classic denoising method and the contrast enhancement method, the proposed method can fill and repair the backscattered light area. The peak signal to noise ratio obtained by the proposed method increases by an average of 9.10 dB and the feature similarity increases by an average of 0.11. The denoising, enhancing contrast, improving the non-uniform illumination are achieved and the backscattered light is removed well.