Subjective and Objective Quality Assessment of Multi-Attribute Retouched Face Images

Facial retouching, aiming at enhancing an individual’s appearance digitally, has become popular in many parts of human life, such as personal entertainment, commercial advertising, etc. However, excessive use of facial retouching can affect public aesthetic values and accordingly induce issues of mental health. There is a growing need for comprehensive quality assessment of Retouched Face (RF) images. This paper aims to advance this topic from both subjective and objective studies. Firstly, we generate 2,500 RF images by retouching 250 high-quality face images from multiple attributes (i.e., eyes, nose, mouth, and facial shape) with different photo-editing tools. After that, we carry out a series of subjective experiments to evaluate the quality of multi-attribute RF images from various perspectives, and construct the Multi-Attribute Retouched Face Database (MARFD) with multi-labels. Secondly, considering that retouching alters the facial morphology, we introduce a multi-task learning based No-Reference (NR) Image Quality Assessment (IQA) method, named MTNet. Specifically, to capture high-level semantic information associated with geometric changes, MTNet treats the alteration degree estimation of retouching attributes as auxiliary tasks for the main task (i.e., the overall quality prediction). In addition, inspired by the perceptual effects of viewing distance, MTNet utilizes a multi-scale data augmentation strategy during network training to help the network better understand the distortions. Experimental results on MARFD show that our MTNet correlates well with subjective ratings and outperforms 16 state-of-the-art NR-IQA methods.