The self-adaptive dwarf mongoose optimized bidirectional network approach for enhanced recognition of overlapped English cursive characters

Automated text recognition systems play a significant role in converting unstructured documents into machine-readable formats, thereby reducing paper usage and creating a paperless environment. Numerous researchers have explored offline handwritten text recognition (HTR) using deep learning methods. The availability of large amounts of data and various algorithmic innovations has made training deep neural networks easier. Extracted features are incorporated to assess the similarity of handwritten characters, which are then classified based on their feature strength. However, these systems are typically limited to recognizing letters, words, and strings due to their high computational cost. To overcome this issue, this paper proposes a novel approach called Bidirectional Network-based Self-Adaptive Dwarf Mongoose (RBN-SADM) for recognizing overlapped English characters with diverse cursive styles. Two different datasets, IAM and EMNIST, containing multiple handwritten samples from various individuals, were used to evaluate the proposed RBN-SADM approach. The collected samples contain various types of noise, such as non-uniform illuminations. Hence, preprocessing steps like thresholding, cropping, normalization, noise elimination, binarization, skeletonization, and tokenization were performed to facilitate recognition. The preprocessed images are trained using the RBN-SADM approach, which accurately learns overlapped features from the images and outputs predicted results with high accuracy. The effectiveness of the proposed RBN-SADM approach was evaluated in terms of f1-measure, accuracy, recall, specificity, error rate, and precision. The analytical results demonstrated that the proposed RBN-SADM approach achieved a higher recognition accuracy rate of approximately 98.7% for the EMNIST dataset and 97.7% for the IAM dataset. Overall, the proposed approach demonstrates significant potential for recognizing overlapped characters with diverse cursive styles.