Dual-Transform Based Noise Estimation

Chongwu Tang, Xiaokang Yang, Guangtao Zhai

Abstract: Noise estimation is the important premise for image denoising. Recently, a noise estimation method based on the scale invariance assumption of high order statistics of natural images' filter response has been proposed and achieved performance superior to classic algorithms. However, in our research, we notice that the scale invariance assumption may not hold well for images with highly directional edges or with large smooth areas. To rectify the problem and further improve noise estimation accuracy, in this paper, we propose a dual-transform based noise estimation algorithm, in which a cascade of wavelet transform and undirectional DCT transform is used. It will be shown that the dual-transform coefficients conforms the scale-invariance assumption well and therefore leads to more accurate noise estimation results. Moreover, computation complexity is considerably reduced due to the down scaling nature of the dual-transform. Extensive experiments and comparative study show the reliability and superiority the proposed methods.