Low-Quality Face Recognition Using an Index-Based Super Resolution Technique

Low-quality face recognition (LQFR), unlike high-quality face recognition is very challenging. This is due to the standoff between the subject and the camera and small face size. The most common way to overcome the mentioned problems is using super-resolution (SR) techniques. In this paper, we show that the efficiency of a FR system degrades significantly in low-resolution images. We propose a novel FR approach using an index-based super-resolution method to improve performance and computational load of a FR system. To achieve this goal, first we apply face quality assessment (FQA) to select appropriate face images with high quality from sequence of images. Then, we introduce Blind/Reference-less Image Spatial Quality Evaluator (BRISQUE) as an index based on which we decided whether or not to use super-resolution in a LQFR system. We also conduct experiments to explore factors which affect FR performance including super-resolution methods, face image size (pixels), face quality and change of the sequence of the conventional LQFR stages. To demonstrate the efficiency of the proposed approach, we compare face identification rate on various deep CNN face recognition. Experimental results show that the proposed method increases the average identification rate to 71.20% on SCFace dataset.