An efficient deep learning system for kinship verification based on ConvNext-EfficientNet-VIT feature fusion

Kinship verification has emerged as a compelling area of research within computer vision, driven by its critical role in real-world applications, such as forensic investigations and the search for missing persons. Despite recent progress, the task remains challenging due to subtle facial similarities across generations and variations in pose, lighting, and expression. Deep learning techniques have significantly advanced the field. Among them, feature fusion has proven to be a powerful tool for enhancing model performance by integrating complementary characteristics from multiple architectures. This research introduces a new kinship verification framework that harnesses the strengths of ConvNext-Base, EfficientNet-B0, and vision transformer (ViT) through an effective feature fusion strategy. By combining the local texture sensitivity of ConvNeXt, the parameter efficiency of EfficientNet-B0, and the global context modeling capabilities of ViT. The fused representation captures a more holistic and discriminative understanding of facial features relevant to kinship. To the best of the authors’ knowledge, this specific fusion of deep models has not yet been explored for kinship verification. The proposed framework is structured into six stages: image preprocessing, parent/child image pairing, feature extraction, feature normalization, feature fusion, and classification. It was evaluated using two standard benchmark datasets – KinFaceW-I (KinFWI) and KinFaceW-II (KinFWII) – achieving maximum accuracy rates of 84.85% and 91.65%, respectively. These results outperform several state-of-the-art (SOTA) methods and underscore the critical role of multi-model feature fusion in improving the accuracy and robustness of kinship verification systems. This research’s promising findings validate the proposed approach’s effectiveness and highlight the potential of deep feature fusion in addressing complex facial analysis problems. • Boosted kinship verification via fusion of ConvNeXt, EfficientNet-B0, and ViT for holistic facial feature learning. • First pixel-wise difference representation for kinship tasks, revealing genetic-linked facial patterns. • Multi-architecture fusion demonstrates superior performance for complex facial relationship analysis.