Image restoration methods for low-resolution and DCT-compressed text images

Restoration from degraded document and video images can generally improve character recognition accuracy. Two common sources that result in image degradation are lossy compression and low-resolution acquisition. Standard image restoration methods, which are designed for grayscale images, often produce inadequate results for text images. The main contributions of this dissertation are to design and develop methods to restore both discrete cosine transform (DCT)-based compressed text images as well as low-resolution text images.; For DCT-compressed images, a Gibbs-Markov random field (GMRF) is introduced to model text as a function of three regions: characters, transition, and background. A set of 27 clique triplets is particularly designed to capture the text characteristics. These triplets are derived from distributions observed in real text images, and used to measure how well each block within the image matches the GMRF-based prior model. In addition to this text model, a constraint is imposed on the DCT-compressed coefficients to preserve the original text image. The restored image is solved iteratively by finding an image that best matches the prior GMRF text model while satisfying the constraint on the given DCT-compressed coefficients. Degraded images were successfully restored experimentally using the proposed GMRF-based model technique.; A novel method is also developed in this dissertation to restore low-resolution text from document and video imagery. The approach uses a Bimodal-Smoothness-Average (BSA) scoring function as an optimal criterion for text image quality. The BSA-based approach is very different from existing methods in the sense that it uses three measures, Bimodal, Smoothness, and Average, to produce a scoring function that represents the quality of a restored image. The idea is to create for a given image a strongly bimodal image with smooth regions in both the foreground and background while allowing for sharp discontinuities at the edges. Document images from the widely available University of Washington database are used for experiments to validate this approach. Optical character recognition accuracy is further used to numerically determine the success of this restoration method. Its superior performance was demonstrated by comparing it to the commonly used linear interpolation and cubic spline expansion techniques.