Text-line Detection and Word Segmentation in Document Images based on an Optimization Framework

Abstract

Locating text-lines and segmenting words in a document image are important processes for various document image processing applications such as optical character recognition, document rectification, layout analysis and document image compression. Thus, there have been a lot of researches in this area, and the segmentation of machine-printed documents scanned by flatbed scanners have been matured to some extent. However, in the case of handwritten documents, it is considered a challenging problem since the features of handwritten document are irregular and diverse depending on a person and his/her language. To address this problem, this dissertation presents new segmentation algorithms which extract text-lines and words from a document image based on a new super-pixel representation method and a new energy minimization framework from its characteristics.

The overview of the proposed algorithms is as follows. First, this dissertation presents a text-line extraction algorithm for handwritten documents based on an energy minimization framework with a new super-pixel representation scheme. In order to deal with the documents in various languages, a language-independent text-line extraction algorithm is developed based on the super-pixel representation with normalized connected components(CCs). Due to this normalization, the proposed method is able to estimate the states of super-pixels for a range of different languages and writing styles. From the estimated states, an energy function is formulated whose minimization yields text-lines. Experimental results show that the proposed method yields the state-of-the-art performance on various handwritten databases.

Second, a preprocessing method of historical documents for text-line detection is presented. Unlike modern handwritten documents, historical documents suffer from various types of degradations. To alleviate these roblems, the preprocessing algorithm including robust binarization and noise removal is introduced in this dissertation. For the robust binarization of historical documents, global and local thresholding binarization methods are combined to deal with various degradations such as stains and fainted characters. Also, the energy minimization framework is modified to fit the characteristics of historical documents. Experimental results on two historical databases show that the proposed preprocessing method with text-line detection algorithm achieves the best detection performance on severely degraded historical documents.

Third, this dissertation presents word segmentation algorithm based on structured learning framework. In this dissertation, the word segmentation problem is formulated as a labeling problem that assigns a label (intra- word/inter-word gap) to each gap between the characters in a given text-line. In order to address the feature irregularities especially on handwritten documents, the word segmentation problem is formulated as a binary quadratic assignment problem that considers pairwise correlations between the gaps as well as the likelihoods of individual gaps based on the proposed text-line extraction results. Even though many parameters are involved in the formulation, all parameters are estimated based on the structured SVM framework so that the proposed method works well regardless of writing styles and written languages without user-defined parameters. Experimental results on ICDAR 2009/2013 handwriting segmentation databases show that proposed method achieves the state-of-the-art performance on Latin-based and Indian languages.