Content Adaptive Hierarchical Parallel Feature Extraction Network for guided thermal image super resolution

Infrared imaging offers robust visual information under low-light conditions, yet its application is frequently hindered by inherently low spatial resolution and degraded textural details. While recent Transformer-based super-resolution models effectively capture global context, their use of fixed window partitioning can limit continuous feature interaction across window boundaries, which may affect the reconstruction of fine structural details. To address these limitations, we propose a Content-Adaptive Hierarchical Parallel Feature Extraction Network for guided thermal image super resolution. In the shallow feature extraction stage, our model employs Dynamic Kernel Convolution (DKC) to adaptively capture initial low-level representations according to local structural variations. For deep feature extraction, we introduce a dual-branch parallel architecture that integrates a Swin Transformer branch with a Varied-Size Window Attention (VSA) branch. The VSA mechanism selects a receptive-field scale (i.e., 3 × 3, 5 × 5, or 7 × 7) for each spatial location according to local variation, enabling spatially adaptive local feature extraction across different structural regions. Furthermore, we propose a Content-Adaptive Spatial Attention Fusion module that applies training-derived weighting maps to coordinate the contributions of the Transformer and VSA features according to local feature characteristics. Extensive experiments on the PBVS benchmark, the MFNet Night dataset, and a custom real-world thermal–visible dataset demonstrate competitive reconstruction performance across diverse acquisition conditions with a compact parameter count of 1.004 million. On the PBVS benchmark, the proposed method achieves the best results in four of the five evaluated metrics, including PSNR, SSIM, MGE, and GMSD, while obtaining the second-best Edge PSNR. However, its relative advantage varies across different acquisition conditions; in particular, RCAN achieves higher PSNR on the Night and real-world datasets. These results indicate that the proposed framework provides a favorable balance between pixel-wise fidelity and structural preservation while effectively utilizing high-resolution RGB guidance. The source code and custom dataset are publicly available at https://github.com/hvxxeo927/CAHPFNet-ThermalSR .