A Hybrid Deep Learning Model for 5-Digit Handwritten Recognition
Alireza Fakhim Babaei, Jafar Tanha, Amir Fakhim Babaei, Mohammadreza Ganbari Matin · 2024
Handwritten digit recognition remains a complex challenge in computer vision, characterized by diverse writing styles, varying distortion levels, and dataset noise. This paper introduces a novel model for predicting handwritten 5 digit images, utilizing Convolutional Neural Networks (CNNs) with an attention mechanism and stacked Gated Recurrent Unit (GRU) networks. We construct a customized dataset by merging MNIST digits and evaluating four model variants. Our top-performing model achieves digit accuracy of 98.73% and sequence accuracy of 93.98%. These results highlight the effectiveness of our approach, emphasizing the role of CNN-based feature extraction with Squeeze and Excitation blocks and stacked GRUs in enhancing sequence prediction accuracy. Additionally, our model offers simplicity, lightweight GRU network design for efficient predictions, and improved accuracy through stacked GRUs, positioning it as a promising solution in the field of handwritten digit recognition.