Deep or Shallow Facial Descriptors? A Case for Facial Attribute Classification and Face Retrieval

With the largely growing quantity of face images in the social networks and media, different face analyzing systems are developed to be employed in real-world situations such as face recognition, facial expression detection, or automated face tagging. Two demanding face-related applications are studied in this paper: facial attribute classification and face image retrieval. The main common issue with most of the attribute classifiers and face retrieval systems is that they fail to perform well under various facial expressions, pose variations, geometrical deformation, and photometric alterations. On one hand, the emerging role of deep CNNs (convolutional neural networks) has shown superior results in tasks like object recognition, face recognition, etc. On the other hand, their applications are yet to be more investigated in facial attribute classification and face retrieval. In this study, we compare the performance of shallow and deep facial descriptors in the two mentioned applications by proposing to exploit distinctive facial features from a very deep pre-trained CNN for attribute classification as well as constructing deep attributedriven feature vectors for face retrieval. According to the results, the higher accuracy of the attribute classifiers and superior performance of the face retrieval system is demonstrated.