Face recognition in low-quality images using adaptive sparse representations

摘要

Although unconstrained face recognition has been widely studied over the recent years, state-of-the-art algorithms still result in an unsatisfactory performance for low-quality images. In this paper, we make two contributions to this field: the first one is the release of a new dataset called 'AR-LQ' that can be used in conjunction with the well-known 'AR' dataset to evaluate face recognition algorithms on blurred and low resolution face images. The proposed dataset contains five new blurred faces (at five different levels, from low to severe blurriness) and five new low-resolution images (at five different levels, from 66 x 48 to 7 x 5 pixels) for each of the hundred subjects of the 'AR' dataset. The new blurred images were acquired by using a DLSR camera with manual focus that takes an out-of-focus photograph of a monitor that displays a sharp face image. In the same way, the low-resolution images were acquired from the monitor by a DLSR at different distances. Thus, an attempt is made to acquire low-quality images that have been degraded by a real degradation process. Our second contribution is an extension of a known face recognition technique based on sparse representations (ASR) that takes into account low-resolution face images. The proposed method, called blur-ASR or bASR, was designed to recognize faces using dictionaries with different levels of blurriness. These were obtained by digitally blurring the training images, and a sharpness metric for matching blurriness between the query image and the dictionaries. These two main adjustments made the algorithm more robust with respect to low-quality images. In our experiments, bASR consistently outperforms other state-of-the-art methods including hand-crafted features, sparse representations, and seven well-known deep learning face recognition techniques with and without super resolution techniques. On average, bASR obtained 88.8% of accuracy, whereas the rest obtained less than 78.4%. (C) 2019 Elsevier B.V. All rights reserved.