Enhancing Handwritten Digit Recognition: A Comparative Study of Basic and Innovative CNN Models Based on the MNIST Dataset

This study compares and analyzes the performance of convolutional neural network (CNN) models combined with residual neural networks (ResNet) and an innovative model in the task of handwritten digit recognition. Experiments were conducted using the MNIST dataset to explore the differences between the two models in terms of accuracy, loss rate, and convergence speed. The basic model, which adopts a classic CNN architecture combined with ResNet, demonstrated efficient training performance and high accuracy. Building on this, the innovative model incorporates residual neural networks, dense connection networks, and attention mechanisms, significantly improving training efficiency and generalization ability. Experimental results show that although the final test accuracy of both models is similar, the innovative model converges faster in the early stages of training and has lower validation loss, demonstrating stronger robustness and stability. This study provides valuable references for further enhancing the performance of handwritten digit recognition models.