Multi-Digit Handwritten Recognition: A CNN-LSTM Hybrid Approach with Wavelet Transforms
Amin Kazempour, Jafar Tanha · 2024
Handwritten digit recognition remains a pivotal area in machine learning and computer vision, essential for applications like license plate identification, form processing, and historical document reading. Addressing the challenges of multi-digit and multi-language recognition, including variations in handwriting styles across different languages, we propose a novel model integrating convolutional and recurrent neural networks with an attention mechanism. Unlike conventional methods, our model employs wavelet transforms instead of max pooling to preserve image texture and edges. We created a comprehensive dataset containing both English and Persian digits, featuring 80,000 training and 20,000 test images with 1–5 digit numbers. To demonstrate the superiority of the proposed model, we conducted extensive experiments and compared it to some state-of-the-art models. Our model demonstrated remarkable accuracy, achieving 99.58% for single digits and 98.03% for sequences. Extensive experiments validated the efficacy of our approach, highlighting its potential for future research in multi-digit recognition systems across various languages.