Cross lingual handwritten character recognition using long short term memory network with aid of elephant herding optimization algorithm

• Published in: Pattern recognition letters 2022-07, Vol.159, p.16-22
• Main Authors: Guptha, Nirmala S, Balamurugan, V., Megharaj, Geetha, Sattar, Khalid Nazim Abdul, Rose, J. Dhiviya
• Format: Article
• Language: English
• Subjects: Enhanced local binary pattern; Gaussian filter; Inverse difference moment normalized; Long short term memory network; Optical character recognition; Projection profile technique
• ISSN: 0167-8655; 1872-7344
• DOI: 10.1016/j.patrec.2022.04.038

•Gaussian filter and skew detection technique are used to improve the data quality.•Projection profile technique is used to segment the line and individual characters.•Used ELBP and IDMN to extract the features from the segmented images.•Used EHO algorithm and LSTM classifier to classify the individual characters. In the recent decades, the handwritten character recognition is still a challenging process in the pattern recognition field. The handwritten digits and characters are not always of the similar width, orientation and size, due to different writing instruments and writing style of the individuals. This makes the handwritten recognition a tricky and tough task. In this manuscript, a new deep learning based model is developed for an automatic character recognition. Initially, the handwritten images are acquired from Chars74K and MADbase digits datasets, and then data pre-processing is carried out using Gaussian filtering and skew detection techniques. Additionally, the individual lines and characters are segmented from the denoised images by utilizing projection profile and thresholding techniques. In addition, the Inverse Difference Moment Normalized (IDMN) and Enhanced Local Binary Pattern (ELBP) descriptors are applied to extract features from the segmented images, and then the discriminative features are selected by employing Elephant Herding Optimization (EHO) algorithm. Lastly, the Long Short Term Memory (LSTM) network utilizes the selected features to classify 64 classes in English language, 10 classes in Arabic language, and 657 classes in Kannada language. Simulation outcomes confirmed that the proposed EHO-LSTM model obtained better performance in handwritten character recognition related to the comparative models. The proposed EHO-LSTM model achieved 96.66%, 96.67%, and 99.93% of accuracy in English, Kannada and Arabic character recognition on chars74K and MADbase digits datasets.