Sitaram Ramachandrula

Latent Dirichlet allocation based writer identification in offline handwriting

In this paper, we describe a novel approach to Writer Identification in Offline handwriting using Latent Dirichlet Allocation. State-of-the-art methods for writer identification employ the traditional feature-classification paradigm which does not provide enough information about the handwriting attributes such as writing style which are key components in any forensic analysis of handwriting. This problem is also compounded due to lack of efficient rules for defining a particular writing style that can capture writer specific characteristics over a large dataset. We propose to address this issue by using a generative model in form of Latent Dirichlet Allocation(LDA) that automatically infers writing styles from handwritten document collection without any pre-defined set of rules. This information is then used to represent each writer as a distribution over multiple writing style for classifying any unknown writer sample. We describe our approach on two different feature sets consisting of contour angle features as well as structural and concavity features. Our experimental results show comparable performance with baseline systems and also demonstrate the efficacy of LDA for learning multiple handwriting styles.

PaperDiff: A Script Independent Automatic Method for Finding the Text Differences Between Two Document Images

In this paper, we introduce a novel concept called {PaperDiff} and propose an algorithm to implement it. The aim of PaperDiff is to compare two printed (paper) documents using their images and determine the differences in terms of text inserted, deleted and substituted between them. This lets an end-user compare two documents which are already printed or even if one of which is printed (the other could be in electronic form such as MS-word *.doc file). The algorithm we have proposed for realizing PaperDiff is based on word image comparison and is even suitable for symbol strings and for any script/language (including multiple scripts) in the documents, where even mature optical character recognition (OCR) technology has had very little success. PaperDiff enables end-users like lawyers, novelists, etc, in comparing new document versions with older versions of them. Our proposed method is suitable even when the formatting of content is different between the two input documents, where the structures of the document images are different (for e.g., differing page widths, page structure etc). An experiment of PaperDiff on single column text documents yielded 99.2 % accuracy while detecting 135 induced differences in 10 pairs of documents.

Using a boosted tree classifier for text segmentation in hand-annotated documents

Highlights? We propose a tree-structured classifier for text identification. ? Increasing classification accuracy for imbalanced data problem. ? Avoid over-fitting for imbalanced set. A boosted tree classifier is proposed to segment machine printed, handwritten and overlapping text from documents with handwritten annotations. Each node of the tree-structured classifier is a binary weak learner. Unlike a standard decision tree (DT) which only considers a subset of training data at each node and is susceptible to over-fitting, we boost the tree using all available training data at each node with different weights. The proposed method is evaluated on a set of machine-printed documents which have been annotated by multiple writers in an office/collaborative environment. The experimental results show that the proposed algorithm outperforms other methods on an imbalanced data set.

Segmentation-free keyword spotting framework using dynamic background model

We propose a segmentation free word spotting framework using Dynamic Background Model. The proposed approach is an extension to our previous work where dynamic background model was introduced and integrated with a segmentation based recognizer for keyword spotting. The dynamic background model uses the local character matching scores and global word level hypotheses scores to separate keywords from non-keywords. We integrate and evaluate this model on Hidden Markov Model (HMM) based segmentation free recognizer which works at line level without any need for word segmentation. We outperform the state of the art line level word spotting system on IAM dataset.

Keyword Spotting Framework Using Dynamic Background Model

An important task in Keyword Spotting in handwritten documents is to separate Keywords from Non Keywords. Very often this is achieved by learning a filler or background model. A common method of building a background model is to allow all possible sequences or transitions of characters. However, due to large variation in handwriting styles, allowing all possible sequences of characters as background might result in an increased false reject. A weak background model could result in high false accept. We propose a novel way of learning the background model dynamically. The approach first used in word spotting in speech uses a feature vector of top K local scores per character and top N global scores of matching hypotheses. A two class classifier is learned on these features to classify between Keyword and Non Keyword.

Offline handwritten word recognition in Hindi

This paper discusses the Hindi offline handwritten word recognizer (HWR) that we are developing. For the purpose of training and testing the offline HWR, we have created a Hindi handwritten word and character database from 100 writers. In our HWR we use two-pass Dynamic Programming algorithm to match the test word against each word in the lexicon by initially segmenting the test word image into probable characters. We extract directional element features (DEF) on each character image segment and statistically model them. Currently we are achieving word recognition accuracies of 91.23% to 79.94% on 10 to 30 vocabulary words.

Real-time embedded skew detection and frame removal

It is common to observe document skew and frame artifacts while photocopying and scanning documents. The motivation of this work is to embed skew correction and frame removal in the copy pipeline of a device to achieve ‘one touch’ cleanup. The two challenges that this poses are the need for: (a) substantially reducing computation and memory requirements and (b) minimizing the false positives. Peripheral document features, such as, page/content edges are low-complexity document skew predictors, and content-based approaches are of relatively higher complexity skew predictors. But state-of-the-art page edge detection methods fail on low-contrast document images, or for similar scanbed/document background. To minimize false positives required in embedded implementations, we propose: (1) a robust page edge detection algorithm that is a multiplicative combination of gradients and line based page edge detectors, (2) a robust skew detection algorithm that is a linear combination of page/content edge and content based predictors, and (3) a pipeline for skew correction and frame removal that uses these algorithms and has near-100% accuracy over a wide range of document images.

Hand-Drawn Symbol Spotting Using Semi-definite Programming Based Sub-graph Matching

In this paper we address the problem of hand-drawn symbol spotting in document images. We use stochastic graphical models (SGMs) to represent the structure and variations of hand-drawn symbols. We use a framework which first carries out segmentation and graph formation of the input image, followed by sub-graph matching for spotting of hand-drawn symbols. We used SGMs in place of sub-graphs in a semi-definite programming based sub-graph matching to do the spotting. The experimental results validate our framework. We were able to spot hand-drawn symbols from 10 classes with 78.89% accuracy in a database of 76 document images and also were able to deal with confusingly similar symbol classes.

Offline handwritten word recognition using MQDF-HMMs

We propose an improved HMM formulation for offline handwriting recognition (HWR). The main contribution of this work is using modified quadratic discriminant function (MQDF) [1] within HMM framework. In an MQDF-HMM the state observation likelihood is calculated by a weighted combination of MQDF likelihoods of individual Gaussians of GMM (Gaussian Mixture Model). The quadratic discriminant function (QDF) of a multivariate Gaussian can be rewritten by avoiding the inverse of covariance matrix by using the Eigen values and Eigen vectors of it. The MQDF is derived from QDF by substituting few of badly estimated lower-most Eigen values by an appropriate constant. The estimation errors of non-dominant Eigen vectors and Eigen values of covariance matrix for which the training data is insufficient can be controlled by this approach. MQDF has been successfully shown to improve the character recognition performance [1]. The usage of MQDF in HMM improves the computation, storage and modeling power of HMM when there is limited training data. We have got encouraging results on offline handwritten character (NIST database) and word recognition in English using MQDF HMMs.

PaperWeb: paper-triggered web interactions

While mobile phones have penetrated deep into tier 2 and 3 cities in India and similar emerging economies, adoption of mobile web content and web services is likely to require the creation of large numbers of relevant applications and services with usable interfaces and interaction paradigms. This paper describes PaperWeb, our effort to enable mobile phone users to use the web for day to day transactions such as paying bills, buying tickets, or fixing appointments, using familiar objects such as paper artifacts. We discuss (i) the creation of useful PaperWeb interactions -- without programming -- by moderately tech-savvy users, and (ii) the use of these interactions by tech-naive users, and briefly describe the underlying technology. We conclude the paper with a discussion of current status and next steps.