In order to address the problems of poor detail resolution and blurred target edges in infrared images, an image enhancement method based on improved Generative Adversarial Network is proposed. Firstly, the generator was constructed based on the codec network U-Net, optimizing the U-Net skip connection method and fusing the global context module to achieve contextual modelling of global and local features. Secondly, the discriminator is constructed based on the Capsule Networks, the capsule network structure is improved by combining with Res2Net structure and the fully connected layer of the Capsule Networks was deconvolutionally reconfigured to achieve multi-scale image feature extraction and reduce model parameters redundancy. The experimental results show that, compared with the current mainstream algorithms, the algorithm in this paper can effectively highlight the detail information, suppress the noise, and improve the image resolution and visual effect.