CNN-Based Image Classification on the Fashion-MNIST Dataset

Abstract

The paper is an empirical study on the use of convolutional neural network (CNN)-based multi-class image classification on the Fashion-MNIST benchmark dataset. Fashion-MNIST, consisting of 70,000 grayscale images of shoes and clothing of ten classes, was chosen as a more difficult follow-up to the canonical MNIST digit recognition benchmark, with significantly higher levels of intra-class and inter-class visual similarity to real world fashion recognition problems. A sequential CNN model was created and trained with TensorFlow and Keras and featured stacked convolutional layers with ReLU activation, max-pooling to perform spatial downsampling, dropout regularization to reduce overfitting, and a fully connected softmax output layer to estimate the probability of classes. The objective function was sparse categorical cross-entropy and the Adam optimizer was used to train the model. The held-out test partition offers empirical assessment with an ultimate classification error of 88.85, which is remarkable generalization and no major indication of overfitting. The high discriminative results of the morphologically distinctive categories, such as trousers, bags, and footwear, and the systematic inter-class confusion observed within the upper-body garment category, specifically the confusion between shirts, T-shirts, and pullovers, indicate that the results are per-class, meaning that the results are influenced by morphologically distinctive categories rather than by arbitrary combinations of such categories. Such results are placed in the context of the larger deep learning literature on fashion image recognition, and the future research directions such as data augmentation, more complex architectures, and attention-based methods are discussed.