Quantization noise is one of the dominant distortions of image compression, and its amplitude usually needs to be estimated for image quality assessment, restoration and enhancement. One such estimation, the peak signal-to-noise ratio (PSNR), has commonly been used as an objective quality measure. However, this measure has limitation in practical applications as it requires as a reference the original image, which is not always available to end users. To overcome the limitation, blind or non-reference PSNR estimation has received much attention in the literature as it requires not the original image, but some statistics of the original image, such as the probability density functions (PDFs) of original discrete cosine transform (DCT) coefficients.
Assuming that PDFs of DCT coefficients follow Laplacian distribution, we propose here a new method to estimate the key parameter of the distribution from a set of training data, consisting of a variety of typical images compressed with various quantization parameters. Our experimental results show that the proposed method can estimate the PSNR of a given image more accurately, with smaller estimation bias and variance, as compared to the existing methods.