Deep One-class Classification: a Deep Learning-based method for people recognition applied to grayscale images

This thesis is aimed at developing an algorithm able to recognize people instances in grayscale pictures. This is a One-Class Classification problem where objects of a particular class are identified compared to all other possible ones. The person class is called positive class or target class, while other items are referred to be in the negative class, also called alien class. The biggest challenge is represented by the variety of objects opposed to the target class, which does neither allow to model the external class in a univocal way, nor to have all possible cases inside the training set. This problem cannot be solved using traditional techniques of binary and multi-class classifications, precisely because there are no pre-defined classes, so no labeled data different from target class objects are available. Examples of the class of interest are categorized just using instances of the same class. The algorithm is built using the Deep One-class Classification (DOC), an innovative Deep Learning-based method targeting OCC problems in computer vision field, like novelty detection, anomaly detection and mobile active authentication. This approach relies on the concept of transfer learning, since an external multi-class dataset from an unrelated task, called reference dataset, is employed to learn deep features characterizing the person class, in addition to the one-class target dataset. A pre-trained Convolutional Neural Network is used in combination with two loss functions, compactness loss and descriptiveness loss, that make the variance of features extracted from the target dataset smaller and minimize the cross-entropy loss of the reference dataset. The proposed approach is able to achieve good results in the Area Under Curve (AUC) of the Receiver Operating Characteristic curve. The classification of people in photo or in video frames makes its way in different areas, like virtual reality, autonomous driving and video surveillance. The use of this algorithm in the latter allows to have big advantages over the user’s privacy, since it works on the device at the tip of the net.