Learning Deep Features for Hierarchical Classification of Mobile Phone Face Datasets in Heterogeneous Environments

In this paper, we propose a convolutional neuralnetwork (CNN) based, scenario-dependent and sensor (mobiledevice) adaptable hierarchical classification framework. Ourproposed framework is designed to automatically categorizeface data captured under various challenging conditions, beforethe FR algorithms (pre-processing, feature extraction andmatching) are used. First, a unique multi-sensor database (usingSamsung S4 Zoom, Nokia 1020, iPhone 5S and Samsung S5phones) is collected containing face images indoors, outdoors,with yaw angle from -90 to +90 and at two different distances,i.e. 1 and 10 meters. To cope with pose variations, face detectionand pose estimation algorithms are used for classifying thefacial images into a frontal or a non-frontal class. Next,our proposed framework is used where tri-level hierarchicalclassification is performed as follows: Level 1, face images areclassified based on phone type; Level 2, face images are furtherclassified into indoor and outdoor images; and finally, Level3 face images are classified into a close (1m) and a far, lowquality, (10m) distance categories respectively. Experimentalresults show that classification accuracy is scenario dependent,reaching from 95 to more than 98% accuracy for level 2and from 90 to more than 99% for level 3 classification. Aset of experiments is performed indicating that, the usage ofdata grouping before the face matching is performed, resultedin a significantly improved rank-1 identification rate whencompared to the original (all vs. all) biometric system.