Evaluation of Contour-based Features for Eyeglasses Style Classification

Eyeglasses style classification is an important task in computer vision with applications in virtual try-on systems, retail, and personalized recommendations. Such systems must operate efficiently in real-time, even under constraints such as limited training data, class imbalance, and variable product imagery. This study explores the effectiveness of contour-based features, specifically Elliptic Fourier Descriptors (EFD), for classifying eyeglass frame styles. EFD coefficients were extracted from three contour perspectives (full frame, half-frame skeletonized, lens boundaries) and combined into a comprehensive feature representation. Classification performance was assessed using various machine learning algorithms evaluated through 5-fold cross-validation on a diverse dataset of frame designs. Nonlinear classifiers, particularly cubic SVM, fine k-NN, and neural networks, achieved superior performance, with validation accuracies exceeding 94%. The findings demonstrate the discriminative capability, robustness, and efficiency of contour-based features, underscoring their potential advantages and practical limitations relative to alternative feature extraction methods.