Classification of Standard FASHION MNIST Dataset Using Deep Learning Based CNN Algorithms
Edmira Xhaferra, Elda Cina, Luçiana Toti
Abstract
Machine learning and deep learning, as one of the most prominent fields of today are quickly improving many aspects of our life. One of the categories that provides strongest results in resolving real-world problems is Convolutional Neural Networks (CNN). Fashion industries have been using Convolutional Neural Networks in e-commerce to solve several problems such as, clothing recognition, clothing searches and recommendations. However, the conventional CNN suffers from several issues including model overfit issues, challenging classification and difficult deep division of garment. It is precisely this complex depth that allows multiple classes to have the same characteristics, making the problem of separation more complex. With this paper, the state-of-art algorithms for the classification of images in the FASHION MNIST database are targeted. Convolutional neural network structures based on deep learning are employed for image classification of the MNIST dataset. The study aims to tackle the model overfit issue, using two different convolutional neural networks CNN-C1 and CNN-C2 architectures to determine which one provides better performance and results. The results show that compared with conventional deep neural network the CNN-C2 outperforms the CNN-C1 architecture and produces higher accuracy of 93.11%.