Performance Analysis of ResNet50 on Balanced and Imbalanced Image Classification Datasets

Abstract

This study explores the performance of a well-known deep learning model ResNet50 on CIFAR-10 and Fashion-MNIST datasets. ResNet50 is an efficient and scalable model for image classification. The model is evaluated on both datasets under balanced and imbalanced conditions. We trained the model using different train-test splits, fine-tuned hyperparameters and applied regularization techniques l1, l2 and dropout. The model is evaluated using Accuracy, Precision, Recall, F1-score, ROC-AUC, Training and Testing time and Memory usage. The results demonstrate that the model performs better on the Fashion-MNIST dataset (86% accuracy) compared to CIFAR-10 dataset (73% accuracy) across all settings. Additionally, the performance of both datasets is better in imbalanced settings but demand more computational resources. These results emphasize how crucial dataset balancing and hyperparameter tuning are to the real-world optimization of deep learning models.