Directionlet Codestream Created By Force Of Interests And Content-based Segmentation

Digital Directionlet Transform (DDT) is a critically sampled and perfect reconstructed technique, it retains the properties of standard two-dimensional (2-D) WT, such as separable filtering, subsampling, the simplicity of computations and filter design. The corresponding anisotropic basis functions (Directionlets) have directional vanishing moments (DVM) along any two directions with rational slopes. Directionlet analysis will benefit image processing in many perspectives.In this thesis, we discuss the existing DDT-based dominant directions determination methods. First, an efficient image segmentation algorithm based on the resemblance of dominant directions in the neighborhood is proposed. Second, a combination of DDT and BFOS is applied for rate consistent coding. Finally, Force of Interests (FOI) priority is introduced to approximate the human visual mechanism in bit allocation. Main contributions include:1. Dominant direction based segmentation is proposed. Segmentations with equal size are performed to obtain an initial partitioning. The dominant directions of each segment are determined to avoid the recursive complexity of Quadtree segmentation. Then, neighboring segments with the same dominant directions are merged. It is efficient to remove the block aritifacts and overcome the Quadtree splitting restriction which images can only be split along the tree nodes.2. To improve the bit allocation efficiency of DDT coding, a modified BFOS algorithm is suggested. For a given bit rate, BFOS algorithm is employed to allocate the appropriate bits for each DDT sub-band. That is, rate-distortion optimization is carried out for codestream truncation.3. Considering the inpacts of ROI in human visual perception, FOI priority adjustment is introduced. By analyzing the sensitivity of edge direction, frequency and ROI, we introduce a priori weight (PW) adjustment for each sub-band and dominant directions of each segment during the bit stream distribution stage. The reconstructed image is more satisfied with human visual characteristics.