Improving Robustness of Image Quality Measurement with Degradation Classification and Machine Learning

Image quality metrics can be classified as generic or degradation specific. Degradation specific measures perform poorly under "mismatched" conditions. Generic measures, on the other hand, may compromise quality measurement accuracy while gaining robustness to variation in distortion conditions. To improve the accuracy-robustness tradeoff, we employ support-vector degradation classification and machine learning tools to judiciously combine generic and degradation specific measures. To test our algorithm, composite quality metrics are optimized for five different distortion classes. Experiment results show that the proposed algorithm achieves improved performance and robustness relative to two benchmark generic quality metrics.

I. Introduction

The most reliable way to measure the quality of images is through the use of subjective quality assessment tests such as the commonly used mean opinion score (MOS) test. These tests, however, are expensive and time consuming, making them unsuitable for automatic quality measurement. Objective (machine-based) measurement methods have been the focus of more recent research. Machine-based measurement allows computer programs to automate image quality measurement in real time, thus playing a crucial role in modern image processing applicationssuch as compression, steganalysis, and communication. Traditionally, error-based quality measures such as peak signal-to-noise ratio (PSNR) and mean squared error (MSE) have been used. Such measures, however, have been shown to correlate poorly with subjective quality scores [1]. Current efforts have focused on devising features that incorporate characteristics of the human visual system.