Transformer-Based Framework for Apparel Image Classification with Data Balancing Method.

Saved in:

Bibliographic Details
Title:	Transformer-Based Framework for Apparel Image Classification with Data Balancing Method.
Authors:	Sebastian, Bob¹ bob.sebastian@binus.ac.id, Kusuma, Gede P.² inegara@binus.edu
Source:	IAENG International Journal of Computer Science. Jun2026, Vol. 53 Issue 6, p2295-2304. 10p.
Subjects:	Transformer models, Data augmentation, Image processing, Image recognition (Computer vision), Deep learning
Abstract:	In recent years, the fashion industry has increasingly embraced deep learning to enhance product identification and inventory management. This study evaluates state-of-the-art vision architectures--ConvNeXt V2, Swin Transformer V2, Data-efficient Image Transformer (DeiT), and BERT Pre-Training Image Transformers (BEiT)--for classifying 59 fashion product categories. To mitigate class imbalance, we applied a comprehensive data balancing strategy combining undersampling and targeted data augmentation. Furthermore, we introduce EdgeNet, a lightweight edge-enhancement module designed to reinforce local structural details often overlooked by transformer models. Among all configurations, the balanced BEiT + EdgeNet model achieved the highest performance, reaching an overall classification accuracy of 83.83%, substantially outperforming other architectures. The results demonstrate that integrating global contextual understanding from vision transformers with localized edge information significantly improves classification accuracy, especially for visually similar items such as loafers and mules or sweat shirts and cardigans. These findings highlight the importance of combining datacentric and model-centric strategies to address fine-grained visual categorization in the fashion domain, offering a robust and scalable solution for real-world applications such as automated tagging, visual search, and personalized recommendation systems. [ABSTRACT FROM AUTHOR]
	Copyright of IAENG International Journal of Computer Science is the property of International Association of Engineers (IAENG) and its content may not be copied or emailed to multiple sites without the copyright holder's express written permission. Additionally, content may not be used with any artificial intelligence tools or machine learning technologies. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)
Database:	Engineering Source

Description
Abstract:	In recent years, the fashion industry has increasingly embraced deep learning to enhance product identification and inventory management. This study evaluates state-of-the-art vision architectures--ConvNeXt V2, Swin Transformer V2, Data-efficient Image Transformer (DeiT), and BERT Pre-Training Image Transformers (BEiT)--for classifying 59 fashion product categories. To mitigate class imbalance, we applied a comprehensive data balancing strategy combining undersampling and targeted data augmentation. Furthermore, we introduce EdgeNet, a lightweight edge-enhancement module designed to reinforce local structural details often overlooked by transformer models. Among all configurations, the balanced BEiT + EdgeNet model achieved the highest performance, reaching an overall classification accuracy of 83.83%, substantially outperforming other architectures. The results demonstrate that integrating global contextual understanding from vision transformers with localized edge information significantly improves classification accuracy, especially for visually similar items such as loafers and mules or sweat shirts and cardigans. These findings highlight the importance of combining datacentric and model-centric strategies to address fine-grained visual categorization in the fashion domain, offering a robust and scalable solution for real-world applications such as automated tagging, visual search, and personalized recommendation systems. [ABSTRACT FROM AUTHOR]
ISSN:	1819656X