CM-TGAN: Twin Generative Adversarial Networks with Convolutional Block Attention Mechanism and Multi-Scale Feature Fusion for Image Dehazing
Dongxu Li, Weizhe Gao, Bingqian Guo, Yikang Yang, Shiping Guo, Jifeng Yan, Ziyang Wang · Advances in transdisciplinary engineering · 2025
Image processing technology is widely applied in many fields, for instance, security and autonomous driving. However, haze significantly degrades image quality. Image dehazing presents significant challenges. In this paper, we proposed CM-TGAN, which means, twin generative adversarial networks (TGAN) with convolutional block attention mechanism (CBAM) and multi-scale feature fusion (MSFF) for Image Dehazing. CBAM is embedded into the networks to improve the capability of capturing key features. MSFF is embedded to enhance the performance of extracting salient information across multiple hierarchical scales. Experimental results show that the proposed algorithm achieves superior performance over TGAN across several critical metrics, for example, peak signal-to-noise ratio(PSNR), structural similarity index (SSIM) and color rendering index (CRI). This algorithm demonstrates significant advantages in edge sharpness and texture detail preservation. Moreover, the proposed algorithm exhibits excellent robustness across diverse illumination conditions and fog concentrations, generating superior-quality reconstructed images that serve as dependable inputs for downstream computer vision applications, including object detection and semantic segmentation.