Using Ensemble Learning to Improve Handwritten Character Recognition Performance

Using Ensemble Learning to Improve Handwritten Character Recognition Performance

This paper proposes a direct and effective approach to improving handwritten character recognition by employing an ensem¬ble learning strategy. By integrating three lightweight convolutional neural network architectures—ResNet, AlexNet, and VGG—through a weighted averaging ensemble, we demonstrate a significant enhancement in prediction accuracy and stability. Even with these relatively simple model structures, the ensemble method effectively harnesses their complementary strengths, leading to improved recognition performance without extensive model optimization. Visualizations from Grad-CAM further illustrate that each base model focuses on distinct features of the input, and through ensemble learning, the combined model mitigates biases and reduc¬es dependency on irrelevant areas such as the background. Our results, validated on the EMNIST dataset, show that the ensemble approach offers a straightforward yet powerful means to enhance both accuracy and robustness in handwritten character recognition tasks, making it particularly suitable for real-world applications where computational efficiency and reliability are also a priority.