Evaluation of a Computer Vision Prototype for Facial Recognition

Abstract

Facial recognition remains a relevant computer vision problem, particularly when illumination, pose, occlusion, and the number of people vary. This study evaluated a prototype developed in MATLAB with Computer Vision Toolbox to detect faces and facial components, store templates, and recognize identities through binary-image correlation. A quantitative, applied, and descriptive approach was used with three sets of ten images: photographs containing one person, two people, and three or more people; an eight-identity gallery with one reference image per subject was also examined. The FrontalFaceCART detector achieved reported effectiveness values of 100% for single-face images, 95% for two-face images, and 74.75% for images with three or more faces. Nose, mouth, and eye detection decreased from 70%, 60%, and 50% in the first scenario to 25%, 37%, and 6.67% in the third. Identity recognition was correct in three of eight cases (37.5%), although one reported success was inconsistent with the 0.30 correlation threshold documented in the code. The results indicate usefulness under frontal and controlled conditions, but limited robustness to crowds, accessories, lighting variation, and sparse reference galleries. Replacing pixel-level comparison with deep representations, validating on standardized databases, and adopting privacy and fairness safeguards are recommended.