Adversarial network for unsupervised infrared image colorization based on full-scale feature fusion and cosine contrastive learning

Thermal infrared images, unaffected by lighting and haze, are widely used in security surveillance, autonomous vehicles, and nighttime traffic monitoring. However, their grayscale nature lacks color and texture details, limiting applications in image recognition and object detection. Converting infrared images to daytime color enhances visual perception and broadens their utility. Despite advancements in infrared image colorization, challenges such as texture distortion, detail blurring, and poor image quality persist. To address these issues, a novel unsupervised learning framework, termed Cosine Contrastive Learning Generative Adversarial Network (CCLGAN), is proposed. Firstly, the traditional UNet architecture is improved by introducing full-scale skip connections and deep supervision. Full-scale skip connections integrate low-level details with high-level semantic features, while deep supervision aids in learning hierarchical feature maps. Additionally, a parameter-free neuron-based 3D attention mechanism is incorporated into the Mamba module to capture long-range dependencies and enable effective feature selection and fusion. Secondly, a novel contrastive loss function is designed, incorporating cosine distance metrics into the traditional contrastive loss framework. By maximizing cosine decision margins and normalizing, intra-class variance is minimized, and inter-class variance is maximized, ensuring consistency between input infrared image patches and output color image patches. Finally, extensive comparative analysis on common datasets demonstrates that the proposed method outperforms existing state-of-the-art techniques in colorization performance. This research advances infrared image processing and enhances the visual quality of converted images. The code is available at https://github.com/LTTdouble/CCLGAN.