Dual Convolutional Neural Network Classifier with Pyramid Attention Network for Image-Based Kinship Verification

• Published in: Acta cybernetica (Szeged) Vol. 26; no. 2; pp. 215 - 241
• Main Authors: Rachmadi, Reza Fuad, Purnama, I Ketut Eddy, Nugroho, Supeno Mardi Susiki, Suprapto, Yoyon Kusnendar
• Format: Journal Article
• Language: English
• Published: Szeged Laszlo Nyul 01.01.2023
• Subjects: Artificial neural networks; Classifiers; Datasets; Neural networks; Verification
• ISSN: 0324-721X; 2676-993X
• DOI: 10.14232/actacyb.296355

A family is the smallest entity that formed the world with specific characteristics. The characteristics of a family are that the member can/may share some similar DNA and leads to similar physical appearances, including similar facial features. This paper proposed a dual convolutional neural network (CNN) with a pyramid attention network for image-based kinship verification problems. The dual CNN classifier is formed by paralleling the FaceNet CNN architecture followed by family-aware features extraction network and three final fully-connected layers. A channel-wise pyramid attention network is added after the last convolutional layers of FaceNet CNN architecture. The family-aware features extraction network is used to learn family-aware features using the SphereFace loss function. The final features used to classify the kin/non-kin pair are joint aggregation features between the pyramid attention features and family-aware features. At the end of the fully connected layer, a softmax loss layer is attached to learn kinship verification via binary classification problems. To analyze the performance of our proposed classifier, we performed experiments heavily on the Family in The Wild (FIW) kinship verification dataset. The FIW kinship verification dataset is the largest dataset for kinship verification currently available. Experiments of the FIW dataset show that our proposed classifier can achieve the highest average accuracy of 68.05% on a single classifier scenario and 68.73% on an ensemble classifier scenario which is comparable with other state-of-the-art methods.