Siamese convolutional neural network and fusion of the best overlapping blocks for kinship verification

Abstract

Analysis of facial images decoding familial features has been attracting the attention of researchers to develop a computerized system interested in determining whether a pair of facial images have a biological kin relationship or not. Given that not all regions of an image are useful to determine the kin relation, thus it is possible to obtain irrelevant and inaccurate information of kinship clues, resulting in false matched kinship. Thus, combining all these regions together will likely produces redundant, irrelevant and deceptive information of kinship, along with higher dimensional space. Motivated by the fact that the facial resemblance among the members in a family can be presented separately in different regions of facial images, where each independent region renders different familial features, there is a high probability that selecting and fusing only the most informative local regions and removing the irrelevant can obtain complementary information for further enhanced accuracy. To this end, unlike other methods, the Fusion of the Best Overlapping Blocks with Siamese Convolutional Neural Network (SCNN-FBOB) is an enhanced method for kinship verification in this paper. This method aimed to simultaneously remove the weak local blocks of the image from a set of overlapping local blocks that achieved low accuracy and only retain the local blocks that achieved high accuracy. Extensive experiments conducted on the benchmark KinFaceW-I and KinFaceW-II databases show highly competitive results over many other state-of-the-art methods.