Transfer Learning-Enhanced Optimization Framework for Handwritten Digit Recognition: A Study of Feature Reusability and Model Fine-Tuning

Abstract

In an era characterized by rapidly increasing data types and volumes, there is a growing demand for machine learning models that can be constructed efficiently while possessing strong generalization capabilities. Transfer learning is particularly well-suited to address these requirements. It is well established that the MNIST dataset consists of handwritten digits collected from foreign populations, and writing styles vary considerably across different countries. Consequently, applying a model trained solely on MNIST to predict handwritten digits from other regions—such as China—often results in suboptimal accuracy in practical applications. To address this limitation, this paper incorporates transfer learning to enhance model performance. First, a custom dataset of the author's own handwritten digit images was collected. Subsequently, transfer learning was applied to a convolutional neural network (CNN) model pre-trained on the MNIST dataset, using this self-collected dataset. Experimental results demonstrate that following transfer learning, model accuracy improved by 7% on the dataset collected by the experimenter.