RDAGAN: Residual Dense Module and Attention-Guided Generative Adversarial Network for infrared image generation

Abstract

Visible-to-Infrared image Translation (V2I) is fundamentally an ill-defined problem, since RGB images do not have any information about the thermal characteristics of different objects. In recent years, with the development of deep learning, infrared image generation has been widely studied, however, existing methods often suffer from the problems of incomplete structure and blurred details in the generated infrared images. For this reason, this paper proposes Residual Dense Module and Attention-Guided Generative Adversarial Networks (RDAGAN) to improve the generation quality of infrared images. RDAGAN incorporates several modules, firstly, we adopt Residual Dense Module (RDM), which improves the model feature extraction capability by enhancing the depth and width of the model. Second, in order to guide the model to focus on the key parts of the image, we designed Attention-Guided Module (AGM), which enable the model to learn and generate the key features of the infrared image more efficiently, thus generating a pseudo-image that is closer to the real infrared image. To further optimize the generated infrared images, we also propose a composite loss function combining the Adversarial loss, L1 loss, Perceptual loss, and SSIM loss, where the Perceptual loss significantly reduces the LPIPS value and improves the visual perceptual quality of the generated images, and the SSIM loss strengthens the edge texture details of the generated images and significantly improves the SSIM value. Experimental results on KAIST, FLIR and LLVIP datasets show that RDAGAN outperforms the existing methods in terms of performance metrics and visual quality, and generates clearer and more realistic infrared images.