A Fusion Schema of Hand-Crafted Feature and Feature Learning for Kinship Verification

The rapid progress of technology is remarkable and becomes more widespread in various forms such as social networks, smart phones, and high-definition cameras. In this context, analysing facial to kinship based on digital images is a new research topic in computer vision and has been increased dramatically in recent years. In this paper, we trying to detect the relationships between pairs of face images which is reflected a verification matter: given a pairs of face images with a view to find out and infer kin from the non-kin. For this, we proposed a method define by a fusion scheme composed of feature learning (high-level feature) and hand-crafted feature (low-level feature) along with features subtracting absolute value for face pair. For hand-crafted, we apply a histogram of oriented gradients (HOG) descriptor, while, convolutional neural net- works (CNN) is to represent the feature learning. In our experiment to validate the proposed method we apply restricted protocol setting. The proposed method is tested and evaluated on the benchmark databases KinFaceW-I and KinFaceW-II, and the verification accuracies of 68.6% and 73.5% were achieved, respectively.