Closing the gap between software and hardware super-resolution image reconstruction: Provision of high-quality output

Abstract

The ability of additional detail extraction offered by the super-resolution image reconstruction (SRIR) algorithms greatly improves the results of the process of spatial images augmentation, leading, where possible, to significant objective image quality enhancement expressed in the increase of peak-signal-to-noise ratio (PSNR). Nevertheless, the ability of providing hardware implementations of fusion SRIR algorithms capable of producing satisfactory output quality with real-time performance is still a challenge. In order to make the hardware implementation feasible a number of trade-offs that compromise the outcome quality are needed. In this work we tackle the problem of high resource requirements by using a non-iterative algorithm that facilitates hardware implementation. The algorithm execution flow is presented and described. The algorithm output quality is measured and compared with competitive solutions including interpolation and iterative SRIR implementations. The tested iterative algorithms use frame-level motion estimation (ME), whereas the proposed algorithm relies on, performance-wise better, block matching ME. The comparison shows that the proposed non-iterative algorithm offers superior output quality for all tested sequences, while promising efficient hardware implementation able to match -at least- the software implementations in terms of outcome quality.