Feature extraction and pattern classification are two basic yet quite important aspects of Pattern Recognition, which outline this thesis. We first discuss the problem of texture image segmentation, and propose a series of principles for filter bank selection to obtain good class separability in the feature space. Then we focus on mean shift, an unsupervised clustering algorithm derived from kernel density estimation. We comprehensively analyze the dynamic mean shift algorithm, and propose the feature space re-sampling scheme for fast mean shift implementation. The primary work of the paper includes: ①We analyze the energy distribution of textures in the frequency domain: its formation mechanism and distribution character, on which a series of adaptive filter bank designing principles are proposed. Orientated to the optimal inter-class reparability, the buildup of the feature space is directly guided by Fisher discriminant criterion, which provides the extracted features with good reparability, low dimensionality and small correlation. With the proposed scheme, satisfactory segmentation result is obtained and the calculation is greatly reduced on the compact filter bank with moderately low dimensions. ②We make further analysis on some basic problems of the iterative clustering algorithm- Mean Shift. (1) We examine the influence of bandwidth selection on sample convergence behavior, and propose a hybrid bandwidth selection scheme which better adapts to sample distribution and achieves higher optimization precision. (2) We strictly prove that the mean shift convergence using Epanechnikov kernel is equivalent to adopting Newton optimization method to locate the local maximum of the density estimated with the same kernel. We also develop a complete Dynamic Mean Shift clustering algorithm based on the deep analysis of the "blurring process" in [51], including the robust bandwidth selection and a new stop criterion. Compared with static mean shift, the convergence of dynamic one is greatly accelerated by the dynamic update of samples and the novelty designed stopping criterion. Theoretical comparisons on convergence between the two algorithms in Gaussian distribution cases, as well as practical clustering and segmentation experiments both prove the efficiency of dynamic mean shift algorithm. ③We combine dynamic mean shift with the spatial information of the image for segmentation tasks, and compare it with static mean shift employed in the joint spatial-range domain. We also develop a feature space re-samplingscheme for fast mean shift implementation. By decomposing the original data set into a series of local sample set with corresponding distribution parameter, the mean shift procedure can be applied in the newly created feature space with much lower computational complexity, therefore greatly speeding up the algorithm. Color image segmentation experiments show that segmentation speed is increased significan
修改评论