Kinship Verification Based On Color Texture Features And Metric Learning

Kinship verification based on facial images is a new topic in the computer vision area,which aims to determine whether a face image pair has a kin relation or not.Kinship verification has a lot of applications,such as missing children finding,social media analysis,family album organization,image annotation et al.At first,kinship verification was researched in the perspective of psychics.Since 2010,machine learning methods have been used in kinship verification.However,existing kinship verification methods based on computer vision have deficiencies in feature extraction and classification.Based on the fact that existing kinship verification methods extract features mainly from gray images which have little information and can be affected by illumination,this thesis proposes an algorithm called kinship verification based on color texture features.First,this thesis takes three color spaces: RGB,YCb Cr and HSV into consideration and apply them in kinship verification.Experimental results indicate that features extracted from three color spaces are better than gray scale features and HSV color space performs best.Moreover,this thesis also discusses a feature extraction by fusing with different color spaces.Experimental results show that fusing HSV with YCb Cr color spaces can get best results.Distance measurement is a key problem in the classification of kinship verification.It is difficult to effectively distinguish between positive and negative pairs in the common use of distance measurement algorithm.In this thesis,the ITML(Information-Theoretic Metric Learning)algorithm is introduced to measure the distance between parents' and children's facial images.By optimizing the distance between image pairs,the distance between positive pairs and negative pairs can be separated by making the distance between positive pairs smaller and distance between negative pairs larger,and the accuracy of verification can be improved.Experimental results show that ITML optimization can improve the accuracy to some extent.Based on above research,this thesis further studies kinship verification with the combination of color features and ITML,and also with the combination of color features and NRML.Experiments illustrate that the combination of color feature extraction and distance metric learning algorithm can further improve the accuracy of kinship verification.