A Transformer-Based Optical Character Recognition Framework with Unified Residual Recurrent Networks for Multilingual Handwritten Documents

Multilingual handwritten Optical Character Recognition (OCR) faces major challenges due to diverse writing styles, script variations, and limited generalization across languages. Existing OCR systems often fail to handle multilingual handwritten data efficiently, resulting in poor segmentation and recognition accuracy. To overcome these challenges, this paper proposes a Transformer-based OCR architecture with a Unified Residual Recurrent Neural Network (TrOCR-URRNN-MLD). The proposed framework integrates Local Sample-Weighted Multiple Kernel Clustering (LSMKC) for effective text segmentation and Adaptive Multi-scale Gaussian Co-occurrence Filtering (AMGCF) for noise suppression and clarity enhancement. A Feature-Affine Residual Network (FA-ResNet) embedded in the Transformer extracts robust spatial-semantic features, while the URRNN with Connectionist Temporal Classification (CTC) efficiently models sequential dependencies. Furthermore, the Secretary Bird Optimization Algorithm (SBOA) optimizes the model parameters for improved performance. Experiments on IAM and Kannada Char74k datasets show that TrOCR-URRNN-MLD surpasses existing OCR models in accuracy, precision, recall, and F1-score.