Date of Award


Document Type


Degree Name

Bachelor of Science (B.S.) in Computer Science and University Honors


Computer Science

First Advisor

Feng Liu


Photographic images -- Rating of, Imaging systems -- Image quality




Existing methods for photo quality assessment typically formulate photo quality assessment as a binary classification problem that labels a photo as low- or high-quality. Photo quality assessment, however, is subjective, and people often rate a photo differently. Therefore, the quality of a photo sometimes cannot be fully described by a low- or high-quality label. In this paper, we present a subjective photo quality assessment method that predicts how a group of users rates a photo. Specifically, our method predicts a quality score distribution that is likely produced by a group of people rating the photo. Our method models the score distribution using the mean and standard deviation. Our method uses a regression approach and integrates a wide spectrum of image features, including manually crafted features, generic image features, and deep learning features, to predict the mean score and standard deviation. We experiment our method on the large scale AVA dataset where each photo on average is rated by 200 users with score ranges from 1-10. Our experiment shows that our regression approach can predict the mean score and standard deviation with RMSE errors 0.67 and 0.19, respectively.

Persistent Identifier