Word Spotting Based on Pyramidal Histogram of Characters Code for Handwritten Text Documents

摘要

In this work, we propose a three-phase convolutional neural networkbased approach for word spotting in handwritten text documents. The system uses a reconstructive convolutional neural network model for segmentation of text document into words. The segmented words are converted into pyramidal histogram of characters code for text representation using modified PHOCNet model. Finally, edit distance is used as a similarity measure between query word and word from text repository. The system is capable of answering the query by example as well as query by a string. The proposed model is also very robust and flexible for the availability of handwritten document repository as training data. The proposed model is validated on IAM dataset of handwritten documents which have 1539 different handwritten text documents from 657 writers.