A multi-dimensional image quality prediction model for user-generated images in social networks

摘要

User-generated images (UGIs) are currently proliferating within social networks. These images contain multi-dimensional data, including the image itself, text and the social links of the owner. UGIs can be utilized for self-presentation, news dissemination and other purposes, and the quality of the image should be able to reveal these social functionalities. However, it is challenging to predict UGI quality utilizing existing models, such as image quality assessment, recommender systems or others, because these models have difficulties processing multi-dimensional data simultaneously. To address this problem, we propose a multi-dimensional image quality prediction model for UGIs in social networks. In this model, we build two sub-models for presentation measurement and distortion measurement. The text (i.e., tags and comments), social links and UGIs are processed by these two models separately, and the results of the models are pooled to obtain a final quality score. Both subjective and objective experiments are then arranged for ground truth data and performance assessment, respectively. Participants are asked to make judgments about 55 UGIs randomly selected from social networks, and the ground truth dataset is based on these subjective experiments. The objective experiments are performed to verify the performance of our model. The results indicate that the Pearson correlation parameter between the predicted score and the ground truth data is 0.5779, which suggests that the proposed model can be implemented to predict image quality in practical environments.