No-reference image quality assessment focusing on human facial region

Abstract A person’s face is the part that we are most interested in when we look at images. If the quality of the human facial regions in focus is poor, even if the quality of other regions is quite better, we are likely to badly evaluate the quality of the image. Thus, to exactly predict perceptual quality of images, the facial regions have to be specially tackled. However, the previous image quality assessment (IQA) methods did not specifically address the facial area. In this paper, we propose a no-reference (NR) IQA metric that uses global features of an image and local features extracted from detected facial regions by a convolutional neural network (CNN) based tiny face detector. Two regressors are used for the global features and the local facial features, respectively. The computed two scores are combined into a total quality score. Promising six IQA databases are utilized for experiments, and it is demonstrated that the performance of the proposed algorithm is competitive with the previous 2D NR IQA metrics.