Comprehensive Feature Extraction Framework for Predicting Perceived Sustainability in Fashion Garments

This research presents a comprehensive feature extraction framework for predicting perceived sustainability in fashion garments. Building upon our prior work that established color as a measurable signal of sustainability perception, this study significantly expands the analytical scope to include texture, shape, and deep learning features. A dataset of 1,080 product images across six garment categories (shirts, T-shirts, jackets, pants, shorts, skirts) was collected from Farfetch and Fashion Product Images Dataset, featuring isolated garments under controlled lighting conditions. A tournament-style selection experiment with 20 participants collected preference data using a composite scoring methodology combining popularity and retention metrics. The framework extracts 342 features per image organized into three categories: color features including colorfulness, brightness, saturation, dominant colors, and earth-tone detection; texture features using GLCM, LBP, Gabor filters, and Tamura descriptors; and shape features via Hu moments, contour analysis, Fourier descriptors, and curvature analysis. Correlation analysis reveals that color entropy (r = −0.297), pattern intensity (r = −0.251), and colorfulness (r = −0.258) are among the strongest predictors, indicating that simpler, less colorful garments are perceived as more sustainable.