Abstract

Ears have been used for the identification of persons since 1910. The first example of how ears were used to recognize humans was the work of an American police officer named Iannarelli. Iannarelli used ear measurements to keep track of prisoners. His works lead to the creation of the discipline known as Biometrics.

Biometric researchers use physical or behavioral characteristics to identify humans. Usually, these physical characteristics are captured in images or video segments. Ears are a valuable resource for biometric researchers performing recognition of humans when a full-frontal image is not available. However, with the advancement of Machine Learning techniques, traditional biometric recognition algorithms such as Local Binary Patterns (LBP), Speeded Up Robust Features (SURF), Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA), and Scale Invariant Feature Transform (SIFT) have proved to be impractical for unconstrained ear segmentation and recognition tasks. This work proposes a Mask-RCNN model for segmentation and a variational Siamese Network Model for recognition and authentication. The project will be tested on a total of 4 different datasets from existing research literature with different lighting, color space, viewpoint, rotation, translation, and yaw. A custom dataset will be created to evaluate the bias of the overall system in terms of skin color, gender, occlusion, and movement in view.