![]() | Only 14 pages are availabe for public view |
Abstract Image denoising is one of the most significant tasks in image processing, analysis and image processing applications. Source signal reconstruction is one of the forms of image denoising, in which we need to retrieve the source signal or image from a noised version, from a mixture with other signals or from a version with missing parts. ICA is a statistical signal processing technique that was originally proposed to find the latent source signals from observed mixture signal without knowing any prior knowledge of the mixing mechanism. The first contribution of our work is proposing a new algorithm SCG-ICA for blind signal separation. The new algorithm significantly improves the convergence rate of gradient-based blind source separation. The proposed algorithm is based on the scaled conjugate gradient method, which used to optimize the kurtosis contrast function in order to estimate the demixing matrix. The second contribution is proposing an ICA based dictionary learning method, which uses the observed data in the learning instead of using fixed dictionary. In order to use the ICA for learning the dictionary, we used the FastICA algorithm under the assumption that the source signals have a Generalized Gamma Density (GGD), which shows a good result in learning the dictionary basis. The results show that learning the dictionary using the ICA are suitable for image denoising and inpainting. We also propose using the Generalized K Distribution (GKD) as an alternative to the Generalized Gamma Density (GGD), and it is also give good results in image denoising. |