A biologically-inspired attentional approach for face recognition

Biometric technologies are now widely used in personal mobile devices. The technological improvements in mobile computing platforms make it possible to embed resource-intensive processes such as human faces recognition. On the other hand, whenever the cooperation of the user is limited, such as in the case of continuous authentication, it is rather difficult to match human performances. The neural architectures and Deep learning has shown great efficiency in the last decade and this mainly due to the computation between the image and the representation in the cortex area. In this paper we propose a biologically-inspired system to perform face recognition by processing image areas captured at different fixation points. The output of simple and complex cells in the V1 striate cortex is simulated by means of a simple convolutional network based on two kind of neurons: S1 and C1. The network layers implement Gabor filters and max pooling operations to encode facial features at different scales and orientations. Higher-level processes, related to face-selective areas, are reproduced through a classification layer. The main inconveniences of the DCNN is the requirement of a huge amount of data. In this paper we propose to add a preprocessing stage to process a small amount of data. The proposed system has been extensively tested against publicly available datasets and the performance has been compared to the current state of the art.