Facial Image Quality Assessment Based on Support Vector Machines

In this paper we propose the first (to the best of our knowledge) overall quality assessment scheme for facial images based on statistical learning. The overall quality assessment system is trained on the subjective quality scores, and is with a high fidelity to the human vision system (HVS) model. This scheme employs a hierarchical binary decision tree classifier based on support vector machines (SVM) to categorize the facial image overall quality into five levels: excellent, good, average, fair and poor. And a classifier fusion process is exploited to improve the performance. In order to train a reliable and generalized system in line with the subjective perception, we construct a large-scale database with 22720 various facial images, which were scored by 10 persons with five quality levels. Experimental results on the database demonstrate that the proposed objective facial image quality assessment system is significantly consistent with the human perception.