Thermal image super-resolution (SR) addresses the limited spatial resolution of infrared sensors, improving the visibility of fine details and supporting applications such as inspection, monitoring, and detection. This capability is particularly valuable in resource-constrained environments where sensor cost, storage, computational resources, and communication bandwidth limit high-resolution thermal imaging. Unlike interpolation, which estimates missing pixels from local image structure, SR aims to recover high-frequency spatial details while mitigating degradations such as blur, noise, and aliasing. This paper presents a comprehensive review of thermal-image SR, covering thermal imaging characteristics, publicly available datasets, evaluation metrics, and representative classical and deep learning methods. It also introduces a demosaicking-inspired deep fusion framework for single-image thermal SR. The proposed approach generates complementary reconstruction candidates using three conventional methods: bilinear interpolation, Malvar–He–Cutler symmetry-based demosaicking, and Menon-based interpolation. A lightweight FusionNet combines these outputs with an RLKA-Net reconstruction by learning spatially adaptive, pixel-wise fusion weights. Experiments on a thermal building-inspection dataset of 306 images demonstrate that FusionNet achieves a PSNR of 37.69 dB and an SSIM of 0.9746, outperforming the individual classical and deep-learning methods. Additionally, a quaternion-based FusionNet reduces model parameters by approximately 75% while maintaining comparable performance, and a higher-capacity quaternion variant slightly surpasses the real-valued model. The paper concludes by discussing remaining challenges and future directions for robust thermal-image SR in real-world applications.