Current Research

• Document Appearance Models: I have been working on building generative models of document appearance in terms of low level simple local features such as Haar filter responses. These features are not robust or stable in the same sense as corner based features (such as SIFT) are designed to be, but are numerous (not sparse) leading to robust distributions. These models are being applied to document image classification and building compositional parts models for document appearance.

  I am also interested in other local image descriptor features for use in such models.

• Functional Role Labeling of parts of document images. The goal of this research is to apply perceptual analysis and machine learning techniques to decompose document images into their functional parts and associate functional labels (such as heading, footer, separators, author blocks, index fields) with these parts.

Previous Projects

• Document Image Decoding: As a member of the the Statistical Pattern and Image Analysis Area (SPIA) at the Palo Alto Research Center (PARC, Inc.), I worked on model directed document image decoding, pattern analysis and content extraction. Detailed models of character templates, and glyph layout models can be trained automatically from images of text, and their transcriptions, and then used for OCR. We obtained 99% or better OCR accuracy on extremely degraded text (see figure), by learning template and layout parameters not just for a specific font but also specific degradation parameters.

  [Papers : font specific training, decoder banks]

• Style consistency in pattern fields (Ph.D. thesis, '00): Exploiting the presence of style/font consistency in recognizing text images. We model the shape consistency of patterns, induced by commonality of origin, on probability distributions in feature space. The theory may be applicable to speech recognition and other applications of statistical pattern recognition.
  Download Ph.D. thesis, or a much shorter paper.

• Human Language Technologies, IBM, T. J. Watson Research Center, Yorktown Heights, NY
  Jun-Dec '98
  Involved in the adaptation of fixed vocabulary speech recognition systems for use in automobiles, and evaluation of speech recognition with various recording devices, under different kinds of noise.

• Panasonic Information & Networking Technology Laboratory, Princeton, NJ
  Jun-Aug '97
  Developed the framework for a software system that identifies logo images in scanned pictures of documents, for the purpose of information retrieval.

• Random phase spatial sampling : Analyzed and modeled random phase effects in spatial sampling of bilevel printed patterns. Theoretical predictions were compared against simulation and experimental results.
  This study has been later applied in designing more robust n-tuple features for OCR.
  In the second phase we developed the concept of modulo-grid diagrams for mathematically analysing the effects of spatial sampling of 2-D bitonal patterns.
  Download paper