Modeling detail feature connections for infrared image enhancement

Ziang Meng, Kaiqi Han, Yuchun He, Yuhan He, Xingyuan Li, Yang Zou

Abstract: While current image super-resolution (SR) methods have shown remarkable achievements across diverse tasks, they often fall short in enhancing infrared (IR) images due to IR-specific characteristics. Notably, the longer wavelengths of infrared radiation, compared to visible light, lead to IR images characterized by lower spatial resolution and less detail. This inherent limitation compromises IR cameras’ capacity to capture high-frequency infrared details. Furthermore, the typical forward propagation in neural networks tends to diminish high-frequency information, making the infrared SR quality more sensitive to the restoration of high-frequency information in infrared images, which existing methods fall short of. To bridge these issues, we propose a long-range sequential processing SR paradigm tailored for infrared images. Specifically, we introduce the Mamba Channel Recursion Module (MCRM) that leverages the selective structured state space model (SSM)’s capability for modeling long-range dependencies to establish a global cross-layer relation that adeptly recovers high-frequency details. Furthermore, we introduce the adaptive deformable self-attention block (AD-SA) and the adaptive deformable cross-attention block (AD-CA) to dynamically offset sampling points, enabling the reconstruction to concentrate more on modal-specific features, and effectively restore infrared high-frequency details. Extensive experiments across various datasets demonstrate that our method notably surpasses existing SR methods, attaining state-of-the-art performance.