Statistical modelling of full frame DCT coefficients

Because of the increasing influence it has in image watermarking applications, the estimation of the distribution shape of full frame DCT coefficients is here addressed. Based on previous analyses on block-DCT, the coefficients are first assumed to follow a Generalized Gaussian distribution. The shape parameter v is then evaluated according to the maximum likelihood criterion applied to a set of 170 natural images. The analysis has been further validated by using the χ² test-of-fit criterion. In contrast to the block-DCT case, experimental results prove that full frame DCT coefficients can be modeled without appreciable loss of performance by a Laplacian density function having variance decreasing with frequency.