Robust Degraded Face Recognition based on Multi-scale Compe- tition and Novel Face Representation

Robust degraded face recognition means the recognizer is robust to low resolution and blurry images and as well as other variations such as illumination, expression and et al. Such task is frequently encountered yet a challenging problem. In this paper, we propose appealing solutions to the task without any image reconstruction and without any blur type limitation. Short-Term Fourier Transform (STFT) is first conducted on face image and then two components relying on STFT are proposed: one is related to window size of STFT named scale and the other is face representation construction from STFT. The goal of the first component is to be robust to low resolution and blur. We propose a multi-scale competition strategy that extracts multiple descriptors corresponding to multiple window sizes of STFT and take the identity corresponding to maximum first candidate confidence as the final recognition result. The goal of the second component is to be robust to other variations. We explore the increased discrimination brought by joint coding and using of multiple frequencies. In particular, we propose a novel local descriptor in which information in local areas coming from two frequencies is jointly encoded and further multiple two-frequency-combinations are jointly utilized so as to construct a more discriminative and descriptive face representation. The experiments conducted on AR and Extended Yale B databases demonstrate that state-of-the-art performance has been achieved by multi-scale competition strategy and the proposed novel face representation.