A novel discrete wavelet transform framework for full reference image quality assessment

In this paper, we present a general framework for computing full reference image quality scores in the discrete wavelet domain using the Haar wavelet. In our framework, quality metrics are categorized as either map-based, which generate a quality (distortion) map to be pooled for the final score, e.g., structural similarity (SSIM), or nonmap-based, which only give a final score, e.g., Peak signal-to-noise ratio (PSNR). For map-based metrics, the proposed framework defines a contrast map in the wavelet domain for pooling the quality maps. We also derive a formula to enable the framework to automatically calculate the appropriate level of wavelet decomposition for error-based metrics at a desired viewing distance. To consider the effect of very fine image details in quality assessment, the proposed method defines a multi-level edge map for each image, which comprises only the most informative image subbands. To clarify the application of the framework in computing quality scores, we give some examples to show how the framework can be applied to improve well-known metrics such as SSIM, visual information fidelity (VIF), PSNR, and absolute difference. The proposed framework presents an excellent tradeoff between accuracy and complexity. We compare the complexity of various algorithms obtained by the framework to the IPP-based H.264 baseline profile encoding using C/C++ implementations. For example, by using the framework, we can compute the VIF at about 5% of the complexity of its original version, but with higher accuracy.