Refining Two Convolutional Neural Networks Architectures for Image Classification Applications

This study explores image classification models using previously constructed models on the Fashion MNIST dataset, containing 70,000 grayscale images divided into ten clothing classes. This study enhances two convolutional neural networks (CNNs) to optimize accuracy, F1-score, and computational efficiency in both training and testing phases. Additionally, it addresses challenges related to overfitting and the classification of multiple classes. Key improvements include data augmentation, dropout layers, modification within the convolutional layers, and class merging. Unlike previous studies that only focused on accuracy improvements, this study balances performance with computational efficiency by systematically modifying CNN architectures. This study is significant in real-world applications requiring precise image classification, such as e-commerce and virtual try-on services. Compared to the baseline model’s accuracy of 90.7%, Model 1 achieved a higher accuracy of 94.7%, while Model 2’s accuracy improved from 89% to 94.1%.