A Reduced Complexity SFSR Upscaler for Embedded Far-Infrared Streaming Video

In this work, Photometric Warp (p-Warp) Single-Frame Superresolution (SFSR) is targeted to an embedded Far-Infrared (FIR) streaming video upscaler system, the goal of which is demonstration of image reconstruction performance and throughput appropriate to mobile or IoT devices that might incorporate infrared imaging technology. In a previous GPU-based implementation, p-Warp SFSR was shown amenable to massively parallel multithreading. However, where an embedded target is considered, thread processing resources remain far less than what might be associated with a modern GPU, one result of which is frame-rate throughput is reduced to an unacceptably low level. In what follows, this problem is addressed via an algorithmic reformulation whereby edge-contour localization and associated warp filter instances are extrapolated to target resolution based upon calculations performed exclusively at source resolution. In this manner, formal time-complexity is reduced and architectural scaling improved. A streaming video upscaler design incorporating this p-Warp SFSR reformulation is then targeted to a mobile i86 platform for which benchmarking tests indicate substantive reconstruction processing gain with increased FPS throughput. Content is organized as follows; (1) Introduction, (2) Photometric Warp (p-Warp) SFSR Overview (3) p-Warp SFSR Algorithmic Reformulation (4) Performance Trial (5) Discussion, (6) Summary.