Research On Kinship Verification Algorithm For Single Sample Based On Sparse Representation And Feature Learning

Facial images with kinship relations contain abundant biological substantive characteristics and associated information. Kinship verification from facial images is a new challenging problem in pattern recognition and computer vision. And the purpose of kinship verification is to determine whether there is a kinship relation between these two given face images.Compared with the traditional method of DNA paternity tests, kinship verification with facial images is more efficient and convenient and so on. This new research topic has several potential applications such as missing children search and video relational analysis. This paper will raise three new algorithms for image recognition and kinship verification based on the analysis and summary of the domestic and international relevant research results.Firstly, a new algorithm based on local similarity pattern matching for single sample was proposed to solve the problem of face alignment under natural scene. On the basis of sub-modular sparse representation algorithms, we add two operations such as modular translation and random assortment to make dictionary atoms and target modulars have the optimum matching as well as multiple features for kinship images to have a better representation. In addition, because voting principles of sub-modular sparse representation algorithms cannot take full advantage of global structure information, we propose a new discriminate regulation that joints the sub-modules voting and global residual to improve the recognition results.Secendly, as for the traditional European metric cannot effectively describe the similarity relation between kinship facial images, we propose neighborhood repulsed metric learning sparse discriminant algorithm which used the similarity of existing data samples to learn a distance metric that can better describe the similarity of samples. Then, the sparse representation method was adopted to establish a dictionary which can be used to linearly represent the children images under the new distance metric space, and the sparse coefficient was applied to measure the similarity of different samples. In addition, to deal with the inconspicuous similarity of kinship samples, a novel algorithm based on sub-modular sparse discriminant was proposed and the multiple sparse coefficients were used to decide whether there is a kinship relation between the two input samples.Finally, the existing feature extraction methods, which always rely on artificial design and have a fixed model, cannot effectively extract a genetic and discriminatory feature. So we propose CNN deep learning model to extract abstract features which can be attained by layer-by-layer eigentransformation for kinship images. The propose of deep learning is to transform the sample feature in the original space to a new feature space so that the new abstract feature is more beneficial to kinship verification and feature-based visualization. In addition, we define a Max-Feature-Map activation function for compact representation and feature selection as an alternative of Re LU.