Thermal image super-resolution (SR) faces critical challenges in recovering fine details from low-resolution inputs due to inherent infrared imaging limitations specifically low contrast between objects, thermal diffusion-induced blurring, and sparse high-frequency components. We propose a computationally efficient dual-input network that leverages frequency-domain analysis with attention-based feature fusion to address these ×2/×4 SR challenges. The architecture first decomposes input images into low- and high-frequency components via Gaussian-blur differencing, explicitly isolating edge information often lost in thermal data. These components undergo parallel processing through a Self-Dual Calibrated Projection Attention (SDCPA) module, enhancing structural coherence while suppressing noise. A multi-scale learning stage then combines three complementary pathways, followed by a Dual Attention Module for dynamic feature fusion before subpixel convolution-based reconstruction. Extensive experiments on multiple thermal datasets (PBVS 2020, FLIR, OSU, CVC09 and Thermal6 datasets) demonstrate that our approach outperforms existing state-of-the-art methods in both ×2 and ×4 upscaling scenarios, achieving superior PSNR and SSIM metrics, while maintaining computational efficiency. This work establishes a practical framework for deploying thermal SR in resource-constrained applications like night vision systems and medical thermography, effectively balancing computational efficiency with high-fidelity reconstruction.