Institutional Repository of Vilnius University
Research on fashion classifier algorithms /
Abstract
dc:descriptionFashion image classification is a critical application of deep learning in e-commerce and retail industries. This research investigates the performance of various Convolutional Neural Network (CNN) architectures for automated fashion classification using the Fashion-MNIST dataset. The study evaluates eight model configurations across five architectures: LeNet-5, AlexNet, ResNet-18, EfficientNet-B0, and Vision Transformer (ViT-B/16), comparing both models trained from scratch and those utilizing transfer learning with ImageNet pre-trained weights. All models were trained on 50,000 images with similar, optimizing hyperparameters using data augmentation and early stopping functions. The results demonstrate that pre-trained EfficientNetB0 achieved the highest accuracy of 95.39%, with precision, recall, and F1-scores of 0.95. Transfer learning consistently improved performance across all architectures. Vision Transformer (ViTB/16) achieved only 88.24% accuracy, highlighting the importance of dataset size for transformer-based architectures. Per-class analysis revealed that simple items such as Trousers and Bags achieved near-perfect classification (𝐹1 ≥ 0.96), while visually similar upper-body garments presented consistent challenges. A functional web application prototype was developed using Flask, enabling users to upload fashion images and receive real-time classification results. This study confirms the effectiveness of CNN-based architecture for fashion classification tasks and provides a practical frameworks prototype for real-world applications.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Jonavičius, Adomas,
Subjects
dc:subject × 1Rights
dc:rights- Statement dc:rights
-
- info:eu-repo/semantics/openAccess
- Language dc:language
- eng
Identifiers
dc:identifier.*- Repository record dc:identifier
- https://repository.vu.lt/VU:ELABAETD277690807&prefLang=en_US