Document image defect models

A lack of explicit quantitative models of imaging defects due to printing, optics, and digitization has retarded progress in some areas of document image analysis, including syntactic and structural approaches. Establishing the essential properties of such models, such as completeness (expressive power) and calibration (closeness of fit to actual image populations) remain open research problems. Work-in-progress towards a parameterized model of local imaging defects is described, together with a variety of motivating theoretical arguments and empirical evidence. A pseudo-random image generator implementing the model has been built. Applications of the generator are described, including a polyfont classifier for ASCII and a single-font classifier for a large alphabet (Tibetan U-Chen), both of which which were constructed with a minimum of manual effort. Image defect models and their associated generators permit a new kind of image database which is explicitly parameterized and indefinitely extensible, alleviating some drawbacks of existing databases.