The infrared camera drifts after two point correction,which will cause non uniform noise in the image.The traditional scene based non uniformity correction algorithm is more adaptable,but it will cause “ghost imaging”.Aiming at the special non uniformity noise of infrared images,the non uniformity correction method based on deep learning will cause image blur,reduced contrast,and loss of details as the network depth increases.A non uniformity correction algorithm based on generative adversarial network is proposed.The network contains two parts:generation network and discriminant network.The generation network use multi scale skip connections to connect each decoder layer and feature maps of different depths,and generate pictures.The discriminant network discriminates the pictures generated by the generation network and the real pictures.Two parts of the network game each other and are trained iteratively.The information of each layer of the network is well preserved through the generation network with multi scale skip connections.After evaluating the effect of infrared image non uniformity correction,it is found that the network can well remove the non uniformity noise of infrared image.The image is clear and the edge is sharp.It retains all kinds of details of the original image and will not produce “ghost imaging”.