Performance Evaluation of Super-Resolution Reconstruction Methods on Real-World Data

Abstract

The performance of a super-resolution (SR) reconstruction method on real-world data is not easy to measure, especially as a ground-truth (GT) is often not available. In this paper, a quantitative performance measure is used, based on triangle orientation discrimination (TOD). The TOD measure, simulating a real-observer task, is capable of determining the performance of a specific SR reconstruction method under varying conditions of the input data. It is shown that the performance of an SR reconstruction method on real-world data can be predicted accurately by measuring its performance on simulated data. This prediction of the performance on real-world data enables the optimization of the complete chain of a vision system; from camera setup and SR reconstruction up to image detection/recognition/identification. Furthermore, different SR reconstruction methods are compared to show that the TOD method is a useful tool to select a specific SR reconstruction method according to the imaging conditions (camera's fill-factor, optical point-spread-function (PSF), signal-to-noise ratio (SNR)).