THE APPLICATIONS OF WAVELET TRANSFORM IN IMAGE PROCESSING

Abstract

Over the past few years, wavelet transform has been receiving a lot of attention m the field of image processing, especially in image and video coding, due to its flexibility in representing nonstationary image signals and its ability to adapt to human vision characteristics. In this thesis, new algorithms are developed for noise removal and multiresolution motion estimation for video compression. The major motivation for applying the wavelet transform for removal of noise from images is the suitability of wavelet transform for finding the location and the spatial distribution of singularities and irregular structures. We proposed the wavelet domain LLMMSE (Local Linear Minimum Mean Square Error) filter for noise removal in Chapter 3. Due to the multiresolution structure of wavelet decomposition, this filter can provide the local adaptation to the human vision system as well as the statistical properties of images at different frequency bands. This filter smoothes noise while preserving edges by taking advantage of different characteristics of image and noise in wavelet domain. A comparison demonstrates that this method provides better improvement of SNR and better subjective impression, especially in detail preservation, than the same method in spatial domain. The multiresolution structure of wavelet decomposition also offers a natural means to implement a coarse-to-fine motion estimation and video compression with spatial scalability. In Chapter 4, a new multiresolution motion estimation scheme based on wavelet transform is proposed. By making use of the motion correlation among different resolution subimages, the motion field in full resolution is obtained by predicting and refining stepwise from coarser resolution to finer resolution until the full resolution is achieved. Due to the shift variance of the wavelet decomposition, the motion estimation is implemented in the smooth subimages. This scheme not only reduces the matching time, but also adapts to the importance of different resolution. In contrast to the conventional full search block matching algorithm implemented in full resolution, this technique yields a good compromise solution between computational complexity and effectiveness of motion compensation. In addition, this scheme outperforms other existing multiresolution motion estimation based on wavelet transform with less computational load and better compensation result. Furthermore, a practical application of this multi resolution motion estimation to a video coding scheme with spatial scalability is given in the same chapter.