Improving Offline Handwritten Text Recognition Accuracy with ADAM and SGD Optimizers and Convolutional Neural Networks Models

Offline Handwritten Text Recognition (OHTR) is a challenging topic in pattern recognition and computer vision due to the variability and complexity of handwriting styles. This study explores the effectiveness of Convolutional Neural Networks (CNN) in improving the accuracy of OHTR systems. Specifically, this study evaluates the performance of ADAM and Stochastic Gradient Descent (SGD) optimizers in training a CNN model. The experiments were conducted on a public handwritten dataset to assess the impact of optimizers on the convergence speed, stability, and overall accuracy of the model. The results show that although both optimizers have advantages, the SGD optimizer performs superiorly in terms of faster convergence and higher accuracy in most scenarios. However, ADAM shows better generalization under certain dataset conditions, indicating that the choice of optimizer may depend on the dataset's characteristics and the specific model's requirements. In this study, it was concluded that using the SGD CNN model gave better results. From 5 experiments with SGD, 5 correct results and 1 incorrect result were obtained, while using ADAM, 2 correct results, and 3 incorrect results were obtained. The researcher realized that things could change based on different CNN conditions and parameters. This study contributes to the development of a more accurate OHTR system by providing a deeper understanding of the effectiveness of optimizers in CNN training. The results of this study are expected to be a guide in choosing the right optimizer for future OHTR systems, with the hope of improving the accuracy and reliability of handwritten text recognition.