Photometric warp-based SFSR with application to infrared image processing
James Glenn-Anderson · 2017
In this work we consider a new Single Frame Superresolution (SFSR) formalism targeting far-infrared video upscaling. In such applications, image reconstruction is rendered more difficult due to Focal Plane Array ('FPA') sensor data of characteristically low resolution, low contrast, and diffraction limited acutance. Despite these challenges, we demonstrate here an SFSR exhibiting accurate and efficient super-Nyquist reconstruction. Theoretical development begins with assumption of a reconstruction principle expressed in form of nonlinear filters exhibiting slope-increasing photometric warp (pWarp) response on edge-contours. These pWarp filters are optimally matched to edge-contour geometry in such manner scale-invariant reconstruction may be achieved. Given any filter instance is by construction specific to an edge geometry it then follows filter application must be conditioned upon a prior structural classification process. We assume for this purpose a tailored differentiable map on image patches by which a manifold emphasizing local edge-contour structure is generated. In the sequel, the resulting pWarp SFSR is shown highly effective when applied to upscaling of a thermal infrared camera data-stream. Content is organized as follows; (1) Introduction, (2) pWarp/SFSR Theory, (3) SFSR Kernel Processing Stack, (4) Performance Benchmark, (5) Discussion, and (6) Summary.