Improved word-level handwritten Indic script identification by integrating small convolutional neural networks

Abstract

Handwritten document recognition has been an active domain of research in the field of computer vision for several years since 1914 with the development of handheld scanner for reading printed texts called “optophone”. In India, which has several different scripts in one document page, identifying them is a must to automate process: document understanding. We propose a novel technique in integrating convolutional neural networks (CNNs) for script identification. We combined small individually trainable small CNNs, and used several different levels of variation in the architectures of the individual CNNs. Such a collection of individually trainable modules vary with respect to the input image size, CNN’s depth and wavelet transformation. In our test, we used publicly available dataset of size 11K words (1K per script) from 11 different Indic Scripts: Bangla, Devanagari, Gujarati, Gurumukhi, Kannada, Malayalam, Oriya, Roman, Tamil, Telugu and Urdu. Several ensemble strategies were implemented such as max-voting and probabilistic voting are used in addition to other conventional approaches like feature concatenation. We achieved a maximum accuracy of 95.04%, and it outperforms the accuracy of the state-of-the-art techniques like AlexNet by 2.9% and more importantly, benchmark techniques as (for script identification) on the dataset by more than 4%.