Optimized Neural Networks for Fashion-MNIST: A High-Accuracy Approach Using Augmentation and Tuning

Image classification is an important subset of machine learning and computer vision that has served as the backbone of various applications in recent years, most notably robotics and autonomous vehicles. The Fashion-MNIST dataset is a collection of 70,000 images widely used as a benchmark dataset. This paper explores three different network architectures with varying complexity. The architectures are run on augmented images from the dataset. The results show that the paper's highest-performing model, ResCNN, achieved a testing accuracy of 98.30%. This surpasses previous CNN-based literature on Fashion-MNIST by 4%, due to the use of residual blocks, a modified architecture, and hyperparameter tuning.