Ph.D. Dissertation Proposal Handwriting Recognition for Document Images Captured by Portable Cameras

The increasing availability of high performance, low priced, portable digital imaging devices has created a tremendous opportunity in document image acquisition by supplementing traditional scanning using flatbed scanners and mounted cameras. However, the portability of the camera presents new challenges in document image analysis and recognition in general and handwriting recognition in particular. Problems are posed by low resolution, blur from focus, uneven lighting and warping distortion. Perspective distortion is one major factor that can cause the recognition accuracy of algorithms designed with the assumption of traditional scanning drop significantly. Features considered robust under the traditional scanning scenario often lose distinguishing power while some features that were not considered before become attractive. In this dissertation, the problems of perspective distortion caused by camera-based document imaging will be systematically studied and new algorithms developed for feature evaluation and classifier construction. We will study theoretically and by implementation the underpinnings of the design of a high performance, lexicon-driven offline handwriting recognizers which can adjust automatically for perspective distortion. Specifically, (i) a new feature evaluation measurement will be introduced to quantify the distinguishing power of features, (ii) a training methodology for automatically learning the distortion parameters is proposed, and (iii) a dynamic feature selection strategy based on perplexity and correlation is proposed to select only a subset of features (from a large set) that exhibit high discriminative power given the automatically computed parameters of perspective distortion. A prototype “perspective distortion independent” recognizer will be built and tested on a dataset of historical document images. First, historical documents are usually fragile and must be subjected to minimal handling during the digitization process. This constraint makes portable digital cameras the digitization device of choice. Secondly, historical documents are predominantly handwritten thus offering a rich source of data to test the algorithms. It is expected that the proposed research will make unique contributions in document image processing, image enhancement, and dynamic feature selection parts of handwriting recognition.