Evaluate the quality of a portrait by analyzing the eyes and the mouth

• Main Authors: Chun-Hui Hsu, 徐君慧
• Other Authors: Jia-Guu Leu
• Format: Others
• Language: zh-TW
• Published: 2009
• Online Access: http://ndltd.ncl.edu.tw/handle/86250955306126020266

碩士 === 國立臺北大學 === 通訊工程研究所 === 97 === With the rapid popularity of digital image capturing devices, which provide a simple way to capture images. Most photos in a daily life are mainly based on full-face photos, however, not every photo of people with good looks. In general, most people agree that details of facial features almost determine whether a photo is good or not. Therefore, we develop an approach and implement a system using the facial features as key information, which is evaluated by the rating system. The rating system can evaluate a full-face photo, we rely on the scores for deciding whether a portrait is good or not, and pick out good portraits we desired from a large amount of digital photos. 　　In this thesis, we first use skin color segmentation to detect a facial area in a photo, this is our face detection stage. Then we also find the accurate coordinates of feature points on eyes and the mouth by using color information. We regard a portrait is good when the eyes are open and not looking side ways, as well as there is a smile. The three features that we defined as a goal to determine whether a photo is good or not, include aspect ratios of eyes, shift between the center of eye and the iris, as well as the curvature on the mouth. Finally, we use the features to develop two rating systems which combines features obtained from the eyes and the mouth. 　　There are 6 persons in our experiments, each person has digital photos that combine 3 different expression in eyes(i.e., normal, half-opened, side-looking) and 4 in mouth(i.e., normal, smile, belly laugh, and curl one's lips downward). In the experiments, we have verified that the experiment results show a good judgment in photo quality. We also rate the portraits manually. We found the correlation coefficients between the manual scores and the scores given by both of our rating systems to be 0.86. Therefore, the evaluation given by our system highly agrees with that of human evaluators.