Research On Image Denoising Based On Multiscale Geometric Transform

Due to the imperfection of image acquisition systems and transmission channels, images are often corrupted by noise. This degradation to a significant reduction of image quality and then makes more difficult to perform high-level vision tasks such as image segmentation and image compression. So how to denoise the image to improve the image quality becomes a very important task in image processing.For the past few years, the basic idea of Multiscale Geometric Analysis(MGA) has developed a series of new theories independently in many areas such as pattern recognition and statistical analysis. This paper focuses on the local directional denosing methods based on the multiscale space. The main work of this paper includes:1. We proposed a novel image denoising method by incorporating the nonsubsampled contourlet transform. The fully shift-invariant property and the high directional sensitivity of the nonsubsampled contourlet transform make the new method a very good choice for image denoising. Firstly, the image was decomposed in different subbands of frequency and orientation responses using the nonsubsampled contourlet transform. Then the multi-scale thresholds were computed according to noise distribution, and used to shrink the nonsubsampled contourlet coefficients. Finally, the modified nonsubsampled contourlet coefficients were transformed back into the original domain to get the denoised image. Simulation results show that the method can obtain higher peak-signal-to-noise ratio.2. Because of the limitation of the threshold, a wavelet-based image denoising using LS-SVM is proposed. Firstly, the noisy image is decomposed into different subbands of frequency and orientation responses using the wavelet transform. Secondly, the feature vector for a pixel in a noisy image is formed by the spatial regularity in wavelet domain, and the LS-SVM model is obtained by training. Then the wavelet coefficients are divided into two classes(noisy coefficients and noise-free ones) by LS-SVM training model. Finally, all noisy wavelet coefficients are relatively well denoised by soft-thresholding method.3. For the past few years, the basic idea of Multiscale Geometric Analysis(MGA) has developed. In this paper, a modified version of the GSM model using Shiftable Complex Directional Pyramid (PDTDFB) is proposed.