Online Bengali handwritten numerals recognition using Deep Autoencoders

Arghya Pal, B. K. Khonglah, S. Mandal, Himakshi Choudhury, S. R.M. Prasanna, H. L. Rufiner, Vineeth N. Balasubramanian

Research output: Chapter in Book/Report/Conference proceedingConference PaperResearch

2 Citations (Scopus)


This work describes the development of online handwritten isolated Bengali numerals using Deep Autoencoder (DA) based on Multilayer perceptron (MLP) [1]. Autoencoders capture the class specific information and the deep version uses many hidden layers and a final classification layer to accomplish this. DA based on MLP uses the MLP training approach for its training. Different configurations of the DA are examined to find the best DA classifier. Then an optimization technique have been adopted to reduce the overall weight space of the DA based on MLP that in turn makes it suitable for a real time application. The performance of the DA based system is compared with systems constructed using Hidden Markov Model (HMM) and Support Vector Machine (SVM). The confusion matrices of DA, HMM and SVM are analyzed in order to make a hybrid numeral recognizer system. It is found that hybrid system gives better performance than each of the individual systems, where the average recognition performances of DA, HMM and SVM systems are 97.74%, 97.5% and 98.14%, respectively and hybrid system gives a performance of 99.18%.

Original languageEnglish
Title of host publication2016 Twenty Second National Conference on Communication (NCC)
EditorsRatnajit Bhattacharjee, Rohit Sinha
Place of PublicationPiscataway NJ USA
PublisherIEEE, Institute of Electrical and Electronics Engineers
Number of pages6
ISBN (Electronic)9781509023615
ISBN (Print)9781509023622
Publication statusPublished - 2016
Externally publishedYes
EventNational Conference Communications 2016 - Guwahati, India
Duration: 4 Mar 20166 Mar 2016
Conference number: 22nd (Proceedings) (Website)


ConferenceNational Conference Communications 2016
Abbreviated titleNCC 2016
Internet address

Cite this