The area of biometrics has seen large advancement in the last decade, although there still remains room for improvement. One of the potentially very useful modalities are ears, which can be used in various different applications in the security and surveillance fields used in combination with other modalities. Ears possess characteristics which can be used to determine the identity of a person relatively well. Therefore, ears can be used in identity recognition tasks. The topic of this master's thesis is to determine whether the characteristics of ears in ear images can be used for kinship analysis. Along with the analysis, a contribution to the thesis is also a dataset consisting of 19 families with ear images for each family member. The images also have manually annotated bounding boxes, which are used for the alignment of the ears, and additionally we provide a list of bad quality images, which can be used to exclude the images from the learning process. For the kinship analysis, a Siamese neural network model is developed, for which 5 different backbones can be used to perform kinship verification via various experiments. Results show that ears are a suitable modality for kinship verification as 4 out of 5 prediction models reach a performance of over 60% of area under the ROC curve, the best performance being 74.1%.
|
