An Hybrid Attention-Based System for the Prediction of Facial Attributes

Recent research on face analysis has demonstrated the richness of information embedded in feature vectors extracted from a deep convolutional neural network. Even though deep learning achieved a very high performance on several challenging visual tasks, such as determining the identity, age, gender and race, it still lacks a well grounded theory which allows to properly understand the processes taking place inside the network layers. Therefore, most of the underlying processes are unknown and not easy to control. On the other hand, the human visual system follows a well understood process in analyzing a scene or an object, such as a face. The direction of the eye gaze is repeatedly directed, through purposively planned saccadic movements, towards salient regions to capture several details. In this paper we propose to capitalize on the knowledge of the saccadic human visual processes to design a system to predict facial attributes embedding a biologically-inspired network architecture, the HMAX. The architecture is tailored to predict attributes with different textural information and conveying different semantic meaning, such as attributes related and unrelated to the subject’s identity. Salient points on the face are extracted from the outputs of the S2 layer of the HMAX architecture and fed to a local texture characterization module based on LBP (Local Binary Pattern). The resulting feature vector is used to perform a binary classification on a set of pre-defined visual attributes. The devised system allows to distill a very informative, yet robust, representation of the imaged faces, allowing to obtain high performance but with a much simpler architecture as compared to a deep convolutional neural network. Several experiments performed on publicly available, challenging, large datasets demonstrate the validity of the proposed approach.