Abstract: Aiming at the low resolution of infrared images, an improved SRGAN super-resolution reconstruction algorithm is designed. In the generative network, the method of applying the residual dense network to obtain the image features extracted from each network layer so as to retain more high-frequency information of the image, and adopting a progressive upsampling method to improve the super-resolution reconstruction effect under a large scaling factor. In terms of the loss function, the perceptual loss that is more in line with human senses is adopted to make the generated image being closer to the real high-resolution image of senses and content. Experimental results show that the quality of reconstructed infrared image is better than that of the current representative methods in the subjective and objective evaluation.

Key words: infrared image, super-resolution reconstruction, generative adversarial network, residual dense network