Stationary wavelet transform features for kinship verification in childhood images

Estimating blood relationships from facial images is a complex task in computer vision due to subtle differences. Existing models are facing difficulties in accurately assessing kin relationships, particularly when dealing with larger age variation datasets. This article proposes a novel approach that addresses these challenges and three key contributions. Firstly, childhood images are introduced as a new dimension for kinship verification, enabling the model to capture relevant facial features despite significant age discrepancies. Secondly, the extraction of facial features from low-illuminated images is a challenging problem. To overcome this, the proposed method employs stationary wavelet transformed (SWT) features and discrete cosine transformed (DCT) features. These features are enhanced through local patch-based SWT (LP-SWT) and LP-SWT-DCT methods. Basic statistical operations, such as mean and standard deviation, are applied to each patch for effective feature correlation. Finally, a selective variance-based method (SVBM) is proposed for metric learning. It selects distinguishable kin and non-kin pairs to design a global threshold, minimizing misclassifications. The proposed model achieves high accuracies: 85% for KinFaceW-I, 89% for KinFaceW-II, 88.90% for UBKINFACE, 85.11% for TSKINFACE, and 88.08% for KFVW. It surpasses state-of-the-art schemes in efficacy.