| This dissertation addresses problems related to efficient structures and design of digital filter banks and proposes a variety of solutions together with applications to image analysis. A large part of the research effort is devoted to novel configurations, design, and efficient implementation of two-dimensional directional filter banks, and their application to processing large aerial images and optical coherence tomography images for isolating directional information. Additionally, a flexible configuration based on phase approximation for one-dimensional M-channel filter banks is proposed, and design techniques that factors in the structural constraints are described.;Efficient structures for image analysis using a large set of two-dimensional directional filters of very large orders with highly narrowband supports configured from one-dimensional building blocks are proposed. Structures of filters for some orientations are obtained by upsampling or downsampling of available impulse responses, where appropriate polyphase representations are exploited to retain the property of generalized separability that allow more efficient implementation when compared with conventional two-dimensional implementation.;The research examined flexible structures for one-dimensional M-channel filter banks based on lifting, which led to a novel filter configuration based on phase approximation. The procedure provides insight into the lifting scheme by interpreting lifting components ideally as fractional delay elements. The prediction filter coefficients used in the lifting framework are designed using an iterative algorithm based on the method of projections on convex sets.;The research also demonstrates the effectiveness of these filters in the analysis of airborne surveillance images and optical coherence tomography images. The directional filter banks are applied to very large aerial images containing small targets with strongly oriented edges; the efficiency of the method compared with conventional implementation is established through simulations. An automated procedure has been developed in which directional filtering is applied to optical coherence tomography images of the retina followed by image segmentation to identify six retinal layers. Results of our investigation show that the error variance in layer thickness estimates using the proposed method is comparable to the variance of differences in expert segmentation. |
