Handwritten character recognition using skewed line segmentation method and long short term memory network

In recent decades, character recognition gained more attention among the researchers, due to its rapid growth in intelligent mobile terminals and communication technologies. Still, handwritten character recognition is a complex task, because the characters differ from writing style, shapes and writing device. To address the aforementioned problems, a novel hybrid model is proposed in this paper to improve the performance of handwritten character recognition, especially for Kannada Arabic and English languages. Firstly, handwritten images are collected from Chars74K dataset, MADbase digits dataset, and a real time Kannada handwritten dataset. Additionally, skewed line segmentation method is developed to segment the individual characters from acquired handwritten image. Then, hybrid feature extraction is accomplished utilizing steerable pyramid transform and discrete wavelet transform to extract the discriminative feature vectors from the segmented images. The extracted feature vectors are given as the input to long short term memory for classifying the individual characters. The simulation analysis showed that the proposed model achieved 95.06%, 99.89% and 98.78% of recognition accuracy in Kannada, Arabic and English handwritten character recognition on Chars74K and MADbase digits dataset. Additionally, the proposed model obtained 88.8% of accuracy in Kannada handwritten character recognition on a real time dataset. Obtained experimental result is better compared to the prior models: consecutive convolutional activations, dentate gyrus of hippocampus, deep contextual stroke pooling, context aware model, and adapted deep hybrid transfer model.