A Comparative Study of ResNet Depth and Training Strategies
Yinuo Sheng · Applied and Computational Engineering · 2025
Deep Convolutional Neural Networks (DCNNs) have become the cornerstone of image classification tasks. Among them, Residual Network (ResNet), as a landmark discovery, occupies an irreplaceable position in both academic research and practical applications. Although many excellent architectures based on ResNet have emerged, the techniques accumulated through modern practices can also significantly improve model performance without changing the ResNet model structure itself. This study investigates the impact of network depth and training strategies on ResNet performance, comparing the performance of ResNet-18 and ResNet-34 on ImageNet subsets and CIFAR-10. It also evaluates the impact of image augmentation and different learning rate scheduling strategies on training outcomes. The experimental results show that deeper ResNet models achieve superior performance, although careful adjustments are needed to mitigate overfitting, especially on smaller datasets. In addition, combining image augmentation can significantly improve the generalization ability of the model. Among the learning rate schedulers tested, Cosine Annealing provided the most stable training process and yielded the best final performance. This work concludes that while increasing model depth can improve representation capacity, employing effective training techniques such as data augmentation and appropriate learning rate scheduling is crucial for optimizing model performance without altering the network architecture itself, highlighting their practical importance in computer vision tasks.