Word spotting: a new approach to indexing handwriting

Abstract

There are many historical manuscripts written in a single hand which it would be useful to index. Examples include the early Presidential papers at the Library of Congress and the collected works of W. B. DuBois at the library of the University of Massachusetts. The standard technique for indexing documents is to scan them in, convert them to machine readable form (ASCII) using Optical Character Recognition (OCR) and then index them using a text retrieval engine. However, OCR does not work well on handwriting. Here an alternative scheme is proposed for indexing such texts. Each page of the document is segmented into words. The images of the words are then matched against each other to create equivalence classes (each equivalence classes contains multiple instances of the same word). The user then provides ASCII equivalents for say the top 2000 equivalence classes.The current paper deals with the matching aspects of this process. Due to variations in even a single person's handwriting, it is expected that the matching will be the most di cult step in the whole process. Two di erent techniques for matching words are discussed. The rst method, based on Euclidean distance mapping, matches words assuming that the transformation between the words may be modelled by a translation (shift). The second method, based on an algorithm developed by Scott and Longuet Higgins, matches words assuming that the transformation between the words may be modelled by an a ne transform.Experiments are shown demonstrating the feasibility of the approach for indexing handwriting.

Fisher Kernels for Handwritten Word-spotting

“…Handwritten word-spotting (HWS) is the task which consists in detecting keywords in collections of handwritten document images [6]. This is a typical two-class pattern classification problem: given a candidate word image and a keyword hypothesis, a match is declared if the score of the word image on the keyword model exceeds a threshold.…”

“…While early approaches to QBE HWS were based on holistic features [6,4], the state-of-the-art is to characterize a word-image with a sequence of feature vectors extracted using a sliding window approach. In their influential work, Rath and Manmatha proposed to use the dynamic time warping (DTW) algorithm to match word images [9].…”

Handwritten Word Spotting with Corrected Attributes

“…Early examples of holistic representations are the works of Manmatha et al [18] and Keaton et al [14]. In [18], a distance between binary word images is defined based on the result of XORing the images.…”

“…In [18], a distance between binary word images is defined based on the result of XORing the images. In [14], a set of features based on projections and profiles is extracted and used to compare the images.…”

Graph-Based Keyword Spotting in Historical Handwritten Documents

“…Keyword Spotting (KWS) is the task of retrieving any instance of a given query word in speech recordings or text images [1][2][3]. Textual KWS can be roughly divided into online and offline KWS.…”