Design of Naturally Distorted Image Database-NDID

The rich content of the electronic images challenge the existing image quality assessment (IQA) models and image aesthetics assessment models. In the meanwhile, it is difficult to get undistorted images, which is important in traditional databases. Therefore, establishing new database containing naturally distorted images is in urgent need. This database contains both image quality factors and image aesthetics factors, which can meet the future research needs of these two types of models. In this paper, we completed the following tasks majorly: (1) We collected 807 naturally distorted images of four semantics in three ways. (2) A subjective evaluation experiment was designed to allow subjective evaluators to rate and label the distorted images comprehensively. (3) We apply NDID to various kinds of image evaluation models and popular deep learning networks. As experimental results showing, NDID have achieved a flat or slightly lower performance compared with other existing databases used in these models. This even or poor performance indicates that NDID has certain practical value for the improvement of images evaluation models.