Enhanced Pix2Pix Visible-to-Infrared Image Generation Network

To solve the problems of inaccurate structural semantic information and poor detail quality of the simulated infrared images generated by existing visible-to-infrared image generation networks, an enhanced Pix2Pix visible-to-infrared image generation network is proposed. Firstly, a Spatial Channel Attention Feature Fusion (SCAFF) module is proposed. This module improves the fusion quality of the high-level features and low-level features of the decoder in the generator, enabling the fusion result to possess more accurate structural semantic information and effectively enhancing the quality of the simulated infrared images. Secondly, a detail enhancer is proposed. This enhancer embeds the features of visible images at different scales into the generation results through feature pyramid pooling, effectively improving the detail quality of the simulated infrared images. The experimental results on the VEDAI, KAIST, and CVC datasets show that the generation results of our network are superior to those of the existing networks in both objective evaluation metrics and subjective visual effects.