Akademik Veri Yönetim Sistemi
5. INTERNATIONAL GAZİANTEP SCIENTIFIC RESEARCH CONGRESS, 27 - 28 Aralık 2025, ss.417-424, (Tam Metin Bildiri)
This study investigates the extent to which classical image processing methods can compete with existing deep learning approaches on a small-scale human face image dataset. Experiments were conducted on the ORL face dataset, and classification was performed using Support Vector Machines (SVM) with Local Binary Patterns (LBP) and Histogram of Oriented Gradients (HOG) features extracted as part of the classical methods. For comparison, a deep learning approach was considered, and deep representations obtained using only the MobileNetV2 model as a feature extractor were classified using SVM. The experimental results show that the classical HOG-based approach offers high and stable performance (92.5% accuracy) under small data conditions and produces results very close to the deep learningbased reference method. Although the MobileNetV2-based hybrid approach achieved the highest accuracy (95%), the classical method's significantly lower computational cost and similar performance without external data dependency are noteworthy for small-scale applications. The findings reveal that classical image processing methods still offer a valid and practical alternative in small data scenarios.