Our project is set to pioneer the classification and grading of clothing on flat surfaces, distinguishing it from most existing classification systems that identify clothing worn by humans. Recognizing the challenge that differentiating between wrinkles and occlusions poses to the grading of clothing, we propose a novel learning-based method. This method is centered on utilizing 3D mesh representations of clothing to accurately identify wrinkles, thereby overcoming a significant hurdle in textile grading. Unlike current systems that excel in identifying clear-cut defects like color fading and fabric tears, our project focuses on the nuanced task of distinguishing minor fabric anomalies, such as wrinkles from tears, which are essential for precise textile grading and effective waste management.
To bridge this gap, we introduce an Automated Wrinkle Detection System that employs deep learning and self-supervised learning techniques. This system is specifically designed to discern wrinkles in fabrics, aiming for a high degree of accuracy in classifying minor textile defects. By developing a model that undergoes a proxy task designed to learn the detailed representation of clothing—such as predicting clothing rotation or identifying edge filters—our approach enables the model to grasp the subtle distinctions in fabric surfaces crucial for detecting wrinkles. We will leverage convolutional neural network (CNN) architectures like ResNet, EfficientNet, and VGGNet, chosen for their respective strengths in depth, efficiency, and detail capture, to ensure our system not only fills the existing technological void in textile grading but also facilitates more sustainable waste management practices by enhancing the accuracy of textile sorting and recycling. Data and Code Available upon request