CS Postdoc Profiles

Research Profile

With advances in digitization of document images and growth of several mass digitization projects, document image analysis and retrieval have emerged as key research areas actively pursued by a number of government agencies as well as industrial research labs. In terms of its scientific impact, this area lies at the unique intersection of machine learning, computer vision and natural language processing and has been continually advancing state-of-the-art in these areas as well as benefitting from their novel research. My research in this area has focused on developing practical solutions that can be applied on the real-world datasets. I am especially interested in adapting ideas from machine learning and information retrieval which provide a strong probabilistic framework for modeling various complex aspects of the problem into a generic solution space.

My dissertation focuses on providing efficient digital access to handwritten document collections. Although research advances in information retrieval has improved the quality of document search to a large extent, such methods are mostly applicable to documents with ASCII text and machine printed document images where the task of automatic word recognition is easier. On the other hand, recognition of unconstrained handwritten documents is difficult due to a large variation in writing styles, document image quality and a large vocabulary which severely degrades the performance of information retrieval techniques .

My approach to this problem is three pronged which derives solutions from machine learning, computer vision and information retrieval. The motivation is to apply learning models which can represent uncertainties associated with noisy text. The other guiding factor is to use as much information as possible while determining the relevance of document images with respect to an input query. In our case, the new sources of information can be found in the document image characteristics or meta-descriptors (e.g. authorship, document class) which are robust to the OCR errors. Finally, an aggregation of all the information extracted at various levels would aim to discount the errors introduced by the recognition phase and provide a robust, federated view of the data for efficient document search.

Contact Information

E-Mail: EMAIL OBFUSCATED

Phone: 7165105752

Categories Posted To:

AI / Machine Learning / Robotics / Vision, Databases / Information Retrieval / Data Mining