Fusion features ensembling models using Siamese convolutional neural network for kinship verification

Abstract

Family is one of the most important entities in the community. Mining the genetic information through facial images is increasingly being utilized in wide range of real-world applications to facilitate family members tracing and kinship analysis to become remarkably easy, inexpensive, and fast as compared to the procedure of profiling Deoxyribonucleic acid (DNA). However, the opportunities of building reliable models for kinship recognition are still suffering from the insufficient determination of the familial features, unstable reference cues of kinship, and the genetic influence factors of family features. This research proposes enhanced methods for extracting and selecting the effective familial features that could provide evidences of kinship leading to improve the kinship verification accuracy through visual facial images. First, the Convolutional Neural Network based on Optimized Local Raw Pixels Similarity Representation (OLRPSR) method is developed to improve the accuracy performance by generating a new matrix representation in order to remove irrelevant information. Second, the Siamese Convolutional Neural Network and Fusion of the Best Overlapping Blocks (SCNN-FBOB) is proposed to track and identify the most informative kinship clues features in order to achieve higher accuracy. Third, the Siamese Convolutional Neural Network and Ensembling Models Based on Selecting Best Combination (SCNN-EMSBC) is introduced to overcome the weak performance of the individual image and classifier. To evaluate the performance of the proposed methods, series of experiments are conducted using two popular benchmarking kinship databases; the KinFaceW-I and KinFaceW-II which then are benchmarked against the state-of-art algorithms found in the literature. It is indicated that SCNN-EMSBC method achieves promising results with the average accuracy of 92.42% and 94.80% on KinFaceW-I and KinFaceW-II, respectively. These results significantly improve the kinship verification performance and has outperformed the state-of-art algorithms for visual image-based kinship verification.