Optimized Malayalam Handwritten Character Recognition Model Using a Novel DSC and Stacked Bi-LSTM with Data Augmentation

Handwritten character recognition (HCR) is a field of computer science and artificial intelligence that involves the recognition of handwritten characters and symbols, typically from digital images or scans. Malayalam has many unique letters with complex shapes, resemblance between different characters and cursive nature. Because of this behavior of the language, Malayalam Handwritten Character Recognition (MHCR) is a tough task. Only a few HCR systems can accurately classify all types of Malayalam characters. This study introduces a novel hybrid model named MHCR, which combines an optimized “Depthwise Separable Convolution (DSC)” with “stacked Bidirectional Long Short-Term Memory (Bi-LSTM)”. The factorization process using DSC results in a significant reduction in both model size and computational cost which is novel in HCR field. The integration of these two components definitely improves the overall performance of the model and it significantly reduce the parameter size of the Proposed Model. Throughout the experiment, the proposed model employed a Depthwise Separable Convolution for feature extraction. These extracted features were subsequently classified using “stacked Bidirectional Long Short-Term Memory (Bi-LSTM)” units. Optimization of parameters has been achieved at Feature Extraction level using DSC. High interclass similarities of Malayalam characters which lead to misclassification can be solved by using Bi-LSTM sequence learning model. This contributions will definitely fill the major research gap such as lack of optimization and misclassification identified through literature review. For training and evaluating the model, MNIST dataset as well as an augmented version of self-generated dataset was used. With 99.75% accuracy, the model recognized 90 different classes of Malayalam handwritten characters.