Towards real-time image quality assessment

Abstract

We introduce a real-time implementation and evaluation of a new fast accurate full reference based image quality metric. The popular general image quality metric known as the Structural Similarity Index Metric (SSIM) has been shown to be an effective, efficient and useful, finding many practical and theoretical applications. Recently the authors have proposed an enhanced version of the SSIM algorithm known as the Rotated Gaussian Discrimination Metric (RGDM). This approach uses a Gaussian-like discrimination function to evaluate local contrast and luminance. RGDM was inspired by an exploration of local statistical parameter variations in relation to variation of Mean Opinion Score (MOS) for a range of particular distortion types. In this paper we out-line the salient features of the derivation of RGDM and show how analyses of local statistics of distortion type necessitate variation in discrimination function width. Results on the LIVE image database show tight banding of RGDM metric value when plotted against mean opinion score indicating the usefulness of this metric. We then explore a number of strategies for algorithmic speed-up including the application of Integral Images for patch based computation optimisation, cost reduction for the evaluation of the discrimination function and general loop unrolling. We also employ fast Single Instruction Multiple Data (SIMD) intrinsics and explore data parallel decomposition on a multi-core Intel Processor.