Optimizing deep ensemble learning for ancient Tamil inscription character recognition via metaheuristic weighting

Abstract Character recognition in ancient Tamil inscriptions poses significant challenges due to variations in writing styles, high inter-class similarity, and the complexity of compound characters. This study proposes a novel Weighted Average Deep Ensemble Learning model optimized using the Cuckoo Search Optimization algorithm (Meta-WADE) to address these issues. The model integrates three Deep Convolutional Neural Networks (DCNNs) and two transfer learning models, with their weights optimized via CSO to enhance performance. Extensive experiments on a large dataset of ancient Tamil characters demonstrate the model’s efficacy, achieving state-of-the-art results with an accuracy of 98.85%, precision of 97.11%, recall of 98.90%, and F1 score of 98.00%. Here we show that metaheuristic-based ensemble approaches can significantly improve the accuracy and robustness of ancient character recognition systems.