ISK: Index steering kernel for multi-SWIR-image super-resolution

Abstract

Multi-image super-resolution (SR) techniques reconstruct high-resolution (HR) images from multiple low-resolution (LR) counterparts, yet existing methods often fall short in reproducing high-frequency details when compared to true HR images, especially short-wave infrared (SWIR) images. This discrepancy, coupled with the technological and cost limitations of enhancing high-resolution imaging system performance through lens aperture size and sensor density, impedes the advancement of imaging systems and digital image technology. Addressing this, we propose the Index Steering Kernel (ISK) method, an innovation grounded in Gaussian process multi-image super-resolution, equivariant kernel theory, and steering kernel methods. Validated through comparative resolution chart tests using a SWIR camera (InGaAs detector), our method achieves a TV Line resolution of 60lp/mm in 4-image 2 × SR experiments, closely approaching the 65lp/mm of genuine high-resolution SWIR images and surpassing alternative approaches. In 9-image 3 × SR experiments, ISK consistently outperforms comparative methods. As a non-data-driven approach, ISK incorporates data compression designs, enabling 3 × SR of nine 640 × 512 SWIR images on a single NVIDIA GTX GeForce 1080Ti graphics card without image splitting, proving its computational efficiency while maintaining superior SR quality.