HyTMO-CapBTM: Hybrid Tri-Manner Optimization Enabled Capsule Network and Deep Learning for Overlapping Handwritten Character Recognition

Overlapping handwritten character recognition is a subset of the Handwritten Text Recognition(HTR) framework, wherein the recognition and separation of the individual handwritten characters takes place. Despite the technical advancements, there exist challenges in identifying the overlapped handwritten characters due to the discrepancies in an individual's writing style and similarities in character shapes. Even though enormous deep learning approaches have been developed for overlapping handwritten character recognition, the interpretability and overfitting issues limit their widespread applicability. To address such constraints, the research proposes a Hybrid Tri-Manner Optimization-enabled Capsule Network and Bidirectional long short-term memory (HyTMO-CapBTM) model to recognize overlapping handwritten characters accurately. The incorporation of the Hybrid Tri-Manner Optimization (HyTMO) algorithm with the Capsule Network and Bidirectional long short-term memory (CapBTM) model helps to improve its convergence speed by avoiding local optimal problems. The ensemble primary capsule layer of the proposed model helps to capture spatial relationships between features, which enhances the performance of recognition. The employed stochastic techniques ultimately overcome the major limitations of previous models and correctly recognize overlapped characters. Through experimental analysis, the proposed model achieves a minimum Mean Squared Error of 5.98 and, Root Mean Square Error of 2.45 using the multi-script handwritten signature dataset for 80% training.