Dual decoding generative adversarial networks for infrared image enhancement

Abstract

Infrared imaging technology is vital for security monitoring, industrial detection, and medical diagnosis. However, atmospheric thermal radiation degrades its quality, causing contrast reduction, texture blurring, and non-uniform noise. To address these challenges, this paper introduces a novel infrared image enhancement method using a dual decoding generative adversarial network (2D-GAN). First, internal and external skip connections are designed to enhance high-frequency detail transmission and mitigate gradient vanishing in deep networks, with local details being preserved as a result. Second, a cross-layer attention mechanism is proposed to adaptively adjust feature map weights spatially and across channels, with information loss during encoding-decoding being minimized and texture clarity and structural coherence being improved. Finally, a joint loss function is designed to integrate pixel-level accuracy, semantic consistency, and global structural coherence, with image realism and perceptual quality being enhanced consequently. Experiments demonstrate superior performance over existing methods in comparative and ablation studies on public datasets, confirming excellent enhancement capabilities and generalization.