Study Of Image Deconvolution Theory In Multiscale Transform-domain

In the practical applications, the digital images are always degraded by many blurred and noisy factors. But in many situations, the clear and high-quality images are required, so how to improve the quality of degraded images is an important research. The technology of image deconvolution is the key technology of this research area (such as image restoration, super-resolution reconstruction).At present, the focus of the research of image deconvolution technology is how to overcome the ill-posedness and obtain clear edge and detail information at the same time. Images always exhibit special properties in transform domain, which can't be obtained in spatial domain. These properties can be utilized to get better results than the traditional methods in image deconvolution problem. In this paper, we mainly research and develop three kinds of image deconvolution algorithms in the multi-scale transform domain.1. Image deconvolution algorithms based on wavelet domain statistical model. Aiming at the problems that the results of image deconvolution algorithms based on the wavelet domain hidden Markov tree model is not good enough caused by the unprecise of the model training and the computational efficiency is low, an improved algorithm is proposed. The proposed algorithm employs the modified Fourier-wavelet regularized algorithm to pre-process the degraded image, and generate a sample image, which can be used to be the foundation of the model, the prior knowledge of the original image, and the deconvolution is finished under the MAP framework. In addition, the contextual hidden Markov tree model is adopted to substitute the original model, the description of relationship of the wavelet coefficients in the same scale is reinforced, so the image quality is improved further.2. Research image deconvolution algorithms based on Hopfield neural network, aiming at the poor ability of detail recovery of this kind of algorithms at present, the wavelet domain hidden Markov tree model is introduced into the framework of neural network, and combines the advantages of these two methods, utilize the convergence property of Hopfield network to complete the deconvolution. In this algorithm, the computation of the weight value matrix is a difficult problem.So we propose a method to solve this problem, only the fast Fourier transform and fast wavelet transform are used, and it can be run in parallel on several processor simultaneously, so the efficiency is high.3. When the image is presented by the wavelet transform, the description ability of the directional information in the image is insufficient inherently, it is also a problem in the wavelet domain image deconvolution. So in this paper, based on the research of other wavelet domain deconvolution algorithms, a new algorithm in contourlet domain is proposed. Contourlet transform is a multi-scale and multi-direction image transform, compared with the traditional separatable 2D wavelet transform, contourlet can express the edge and texture of images more efficiently. The proposed algorithm is based on the bounded optimize algorithm, and proceeds in the contourlet domain iteratively. The proposed algorithm is better to recovery the contour and texture of images, compared with the similar algorithms in wavelet domain.