Autoencoder Feature Based Kinship Verification using Machine Learning Classifiers

Abstract

Verification of kinship in an automatic way is a new challenge in vision on the computer that intends to determine if two given individuals have any kinship relationship based on their facial features. Verification of Kinship by facial pictures is difficult since face images have a lot of intra-class variations due to age, genetic and gender difference. Videos can provide more information as they have additional sources of information as compared to single image verification. In this paper, deep learning based Supervised Mixed Norm regularization Autoencoder (SMNAE) is suggested for verification of kinship in unconfined videos. This proposed SMNAE based approach employs the spatial information to test kinship in the video frames. The Random Forest (RF), Support Vector Machine (SVM), and Decision Tree (DT) classifiers are used to Classify the decision borderline among kin and non-kin classes. The efficacy of this proposed approach is evaluated on seven kin relationships using a kinship video (KIVI) face dataset of kin pairs. According to experimental results SMNAE spatial features with DT classifier produce 70.14% of f1score, which performs much better than other classifiers. It gives improvement in f1 score of +24.52%, +22.05% than SVM and RF classifiers respectively. Furthermore, when compared to existing systems, the proposed SMNAE spatial features with DT framework performs exceptionally well in kinship verification.