Blind structural similarity estimation of digital images using quantized discrete cosine transform coefficients

Abstract

Objective image quality assessment is used to develop a quantitative measure in order to predict perceived image quality by exploiting a variety of known properties of the human visual system (HVS). A new paradigm for quality assessment of the image is based on the structural similarity (SSIM) index, which takes advantage of characteristics of the HVS. In order to estimate SSIM, we need to know the source image to quantify the visibility of errors between the distorted image and the referenced image. In many practical applications, however, the reference image is not available and a blind quality assessment should be utilized. In this paper, an algorithm for statistical estimation of the SSIM based on the probability density functions (pdfs) of quantized discrete cosine transform (DCT) coefficients is presented. In the proposed method, we assume that the pdfs of the original DCT coefficients follow a specific distribution. The parameters of this distribution are then obtained from the quantization step size and quantized DCT coefficients of the distorted image, which are then used to calculate the SSIM metric. Our proposed method is, therefore, applicable to the encoding schemes that involve DCT quantization such as JPEG encoding. Numerical results show that the proposed SSIM estimation method provides relative errors that are generally smaller than those of the available peak signal-to-noise ratio estimation schemes for DCT-based images.