Semi-Automatic Geometric Normalization of Profile Faces
Justin Romeo, Thirimachos Bourlai · 2019
This paper proposes a correlation point matching approach, i.e. an efficient methodology for applying geometric normalization for profile face images. This method is used to increase accuracy without imposing a significant increase in face matching computational time when using different feature descriptors. In our work, several such descriptors are tested to compare the accuracy with which low level facial features (edges), useful for profile face image geometric normalization, are extracted. Hence, we determined the most efficient normalization approach that does not substantially increase computational time. Experimental results show that the use of eigenvalues produces a higher than average edge point count, while having a lower increase in computational complexity compared to other similar algorithms. Then, the extracted features are matched using the random sample consensus algorithm (RANSAC). Next, the rotational angles between the pairs of features are calculated and averaged to yield the angle of rotation necessary to achieve a proper profile face image normalization representation. After applying our proposed approach to a deep learning-based profile face recognition algorithm, an increase of 7.2% accuracy is achieved when compared to the baseline (non-normalized profile faces). To the best of our knowledge, this is the first time in the open literature that the impact of automated profile face normalization is being investigated to improve deep learning-based profile face matching performance.