Results: Center for Unified Biometrics and Sensors, State University of New York, Buffalo, USA, and HP Labs India, Bangalore, are working on an approach to segment handwritten text, machine printed text, and noise from annotated machine printed documents. The experimental results on an imbalanced data set has shown that the approach achieves an overall recall of 96.33 percent.

Why it matters: Unlike the retrieval of machine printed documents, where high OCR accuracy can be expected, the retrieval of noisy annotated documents which contain both handwritten text and machine printed text is still a challenge because document retrieval in the context of handwriting has not been widely explored. The Markov Random Fields (MRF) approach paves way for preprocessing mixed documents which is an important step in the design of systems for OCR, author identification, signature verification and document retrieval.

Methods: Initially, three different kinds of patches are classified and then refined using MRF, based on the concept of a system of neighbors and belief propagation. Prior to classification, each binarized document is segmented into patches which are small snippets of the image. In MRF-based framework, each document is modeled as a random field which consists of a number of patches. A specific sized window is used for dilation of the original binarized image and the bounding box of each connected component after dilation is defined as a patch. The size of the window is empirically chosen such that the resultant patch typically represents a handwritten or machine-printed word. Three different categories of features are considered for classification of a given patch into one of three classes: handwritten text, machine printed text, and noise.

Next steps: Experiments show that the MRF method has better classification performance than a single classifier. Future work includes use of smaller patches and use of other classification techniques.