Adaptive Feature Learning in Transfer-Based Visual Garment Classification

Abstract

In the rapidly evolving field of fashion automation and textile manufacturing, accurate identification of garment types and stitching patterns is crucial for quality assurance, inventory management, and automated sorting. With increasing production demands, industries require intelligent systems capable of analyzing large volumes of garment images with high precision and consistency. Traditional methods based on manual inspection or basic image processing lack scalability, consistency, and the ability to detect fine-grained stitching variations under complex conditions. This work addresses the lack of a reliable automated system for dual-level garment classification, specifically identifying both fabric type and stitching condition from images. Conventional approaches rely heavily on human expertise, resulting in inconsistent outcomes, limited accuracy, and inefficiency in real-time applications. Additionally, these methods lack advanced feature extraction and automated learning, leading to poor generalization across diverse datasets. To overcome these challenges, the proposed system, FabricNet, introduces an intelligent framework for automated garment analysis deployed across two systems: Server and Client. The server extracts deep visual features using EfficientNetB0, which are then processed by models such as Deep Neural Network (DNN), Perceptron, and a proposed Multi-Layer Perceptron (MLP) for multi-output classification. A Tkinter-based graphical interface supports administrative tasks including dataset management, preprocessing, model training, and evaluation, while the user interface enables real-time image-based prediction. The system also incorporates explainable AI to interpret predictions. FabricNet enhances accuracy, reduces manual effort, and provides a scalable solution for modern smart garment industries.