Comparison of Filtering Methods for Extracting Transient Facial Wrinkle Features

Facial local features comprise an essential information to identify a personal characteristic such as age, gender, identity and expression. One of the facial local features is a wrinkle. Wrinkle is a small furrow or crease in the skin. Recently, wrinkle detection has become a topic of interest in computer vision, where many researchers developed applications like age estimation, face detection, expression recognition, facial digital beauty and etc. However, most of the research focused on permanent wrinkles instead of transient wrinkles. Transient wrinkle can be seen during the movement of facial muscle such as a facial expression. This paper presents a comparison of filtering method for extracting transient wrinkles features. The filters that have been selected are Gabor wavelet and Kirsch operator. The extracted features are the number of wrinkles, the maximum perimeter of wrinkle, the average perimeter of wrinkle, total perimeter of wrinkle, the maximum area of the wrinkle, and the total area of the wrinkle. A total of 60 sets of data extracted from Cohn-Kanade database, images from internet and self-images. These images contain weak and strong transient wrinkles at forehead region. Features selection and analysis has been done to select which feature extraction method produces better wrinkle features that can be used for the classification of wrinkle detection system. The results show that both Gabor and Kirsch methods are successful to extract transient wrinkle features, where both methods scored 100% accuracy in the classification with SVM. However, Gabor method is slightly better than Kirsch method in term of detecting weak wrinkles. The Kirsch method requires an additional noise filtering method to eliminate noise particles after the convolution of Kirsch’s kernel. In conclusion, Gabor method is more applicable to a variety of applications than Kirsch method.