Proposing a sparse representational based face verification system to run in a shortage of memory

Studying face verification has seen tremendous growth over the past years. During the last decade, with the improvement of system processors and memories, deep learning was growth widely and the applications of Convolutional Neural Network (CNN) affected all image processing tasks. But, needing much space to save several parameters of learned model is still a big challenge to use them on simple devices, e.g. cell phones. In this paper, to address the problem of face verification in a shortage of memory sparse representation has been employed. So, to compare two portraits a dictionary is generated from each image using augmentation techniques. Then, each face is reconstructed sparsely by the other dictionary and if there is a negligible average of reconstruction error, couple of faces are matched. The proposed method has been assessed in various conditions of several face datasets and the results show improvement comparing to all sparse representational approaches. Although the evaluations indicate a bit less accuracy than CNN-based methods, the main advantage is less usage of memory that can lead running on mobile devices.