Ear recognition with ensemble classifiers; A deep learning approach

Biometrics has emerged as a major domain for security systems. Ear as a biometric has many distinctive features which makes it promising for personal identification systems. In this paper, two tracks for classification of ear images are implemented and tested. The first employs a classical machine learning technique based on extracting features from the discrete curvelet transform and passing the extracted features to a classifier. Image preprocessing is needed for enhancement and segmentation. Ear region is first selected from the background then the curvelet transform via wrapping is applied on the segmented ear images. Different levels are investigated. The coarse image is divided into blocks and the mean, variance and entropy are calculated for each block and concatenated with the same calculated statistical features from the subimages at different levels forming the feature vector. The feature vector is passed to a classifier for ear recognition and the only classifier that provided comparative results was the ensemble classifiers. In the second track, deep learning methods are employed. Different end-to-end networks are used for classifying ear images. Features are then extracted from each network and fed to a shallow classifier for ear classification. Principal component analysis is used for feature reduction. Different classifiers are again investigated and the only classifiers which succeeded to give superior results are the Ensemble classifiers. The achieved classification rate showed improved results compared to the published methods that proves the superiority of the Ensemble classifiers for correctly classifying ear images.