MD 2 C-Mamba: Multi-Domain Divide-and-Conquer Mamba Framework for AAV Thermal Image Super-Resolution

Thermal imaging technology significantly enhances the perceptual capability of consumer unmanned aerial vehicles (UAVs) under nighttime, hazy, and other adverse environments. However, the insufficient resolution of thermal images acquired by these systems presents a significant obstacle to autonomous decision-making of UAVs. Furthermore, thermal image super-resolution (SR) encounters challenges such as uniform background pixel distributions and sparse target regions, demanding models capable of effectively capturing long-range dependencies and fine-grained local and global information. In this paper, we propose MD2C-Mamba, a novel divide-and-conquer network based on the Mamba architecture. MD2C-Mamba innovatively parses the latent feature space by considering local and global spatial domains, as well as the frequency domain, enabling progressive refinement of SR results. Specifically, we incorporate a vision-centric state space module (SSM) to capture long-term spatial dependencies, thereby improving the restoration of fine details. Moreover, we design a feature mapping module that leverages wavelet transforms to capture local features and texture information within images, enhancing feature representation through the integration of global and local branches. To address the challenges of modality differences and semantic alignment, we develop a multi-domain attention fusion module (MDAF) that employs cross-attention mechanisms. This enables complementary interaction and fusion of features, enhancing both global and local characteristics and ultimately improving model robustness. Extensive experimental results demonstrate that the proposed MD2C-Mamba method surpasses the performance of state-of-the-art SR methods, yielding more realistic details and achieving superior performance across multiple evaluation metrics.