Multi-Scale Feature Fusion with Attention Mechanism Based on CGAN Network for Infrared Image Colorization

This paper proposes a colorization algorithm for infrared images based on Conditional Generative Adversarial Network (CGAN) with multi-scale feature fusion and attention mechanisms, aiming to address issues such as color leakage unclear semantics in existing image coloring methods. Firstly, we improved the generator of CGAN network by incorporating extraction module into U-Net architecture fuse features from different scales, thereby enhancing network’s ability extract improving its semantic understanding, which improves problems blurriness during colorization. Secondly, enhanced discriminator introducing an mechanism module, includes channel spatial modules, better distinguish between real generated images, clarity resulting images. Finally, jointly both module. We tested our method dataset containing near-infrared retains more detailed while also preserving advantages The experimental results show that proposed achieved peak signal-to-noise ratio (PSNR) 16.5342 dB structural similarity index (SSIM) 0.6385 RGB-NIR (Red, Green, Blue-Near Infrared) testing dataset, representing 5% 13% improvement over original network, respectively. These demonstrate effectiveness addressing network. this is not only applicable but can be widely applied remote sensing CT