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

• Published in: Applied sciences Vol. 13; no. 8; p. 4686
• Main Authors: Ai, Yibo, Liu, Xiaoxi, Zhai, Haoyang, Li, Jie, Liu, Shuangli, An, Huilong, Zhang, Weidong
• Format: Journal Article
• Language: English
• Published: Basel MDPI AG 01.04.2023
• Subjects: Algorithms; attention mechanism module; Colorization; Computed tomography; CT imaging; Datasets; Deep learning; Discriminators; Evaluation; Game theory; Generative Adversarial Network (GAN); image colorization; Infrared imagery; infrared images; Leakage; Medical imaging equipment; Methods; multi-scale feature fusion; Remote sensing; Semantics; Signal to noise ratio
• ISSN: 2076-3417; 2076-3417
• DOI: 10.3390/app13084686

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