Research on the Application of CNN in MNIST Image Classification and Recognition

This paper investigates the application of Convolutional Neural Networks (CNNs) in MNIST handwritten digit recognition, with a particular focus on optimizing the ResNet-18 model. By introducing residual blocks and batch normalization techniques, the problem of vanishing gradients in deep networks was addressed, improving the model's convergence speed and stability. The study systematically fine-tuned two hyperparameters-number of training epochs and batch size-resulting in an optimal configuration that achieved a 99.40% accuracy rate on the MNIST test set. Comparative experiments with classical models such as LeNet-5, AlexNet, and VGG-16 validated the superior performance of ResNet-18 in image classification tasks. The findings of this study not only enhance the accuracy of MNIST image classification but also provide valuable insights for further application to more complex image classification tasks.