Ensemble Learning Based on Convolutional Kernel Networks Features for Kinship Verification
Kinship verification based on facial images is one of the popular research topic in the field of face recognition. It's still a challenging problem due to many inevitable factors, such as varying illumination, poses, and expressions. And traditional handcrafted features are usually not robust enough. For the above reasons, in this paper, we extract kernel features by Convolutional Kernel Networks (CKN), which are invariant to particular transformations. After extracting the CKN features, we use feature bagging to classify. It's an ensemble learning method that attempts to reduce the correlation between estimators in an ensemble by training them on random samples of features instead of the entire feature set. Specifically, the CKN features are randomly sampled to train an SVM classifier each time, and then multiple SVM classifiers are combined by majority voting to make prediction. In addition, we collect a large kinship face dataset named LarG-KinFace from Internet search under uncontrolled conditions. The proposed method is evaluated on three datasets KinFaceW-I, KinFaceW-II, and LarG-KinFace. Experimental results demonstrate the efficacy of the proposed method.