CSIQ Image Quality Database
The CSIQ image database is a popular database for testing image quality assessment algorithms and other aspects of image quality. The database consists of 30 original images, each distorted using one of six types of distortions, each at four to five different levels of distortion. The CSIQ images were subjectively rated based on a linear displacement of the images across four calibrated LCD monitors placed side-by-side with equal viewing distance to the observer. The database contains 5000 subjective ratings from 35 different observers, and the ratings are reported in the form of DMOS (a larger denotes greater visual distortion compared to the corresponding reference image).
If you use this database, please cite the following paper: E. C. Larson and D. M. Chandler, "Most Apparent Distortion: Full-Reference Image Quality Assessment and the Role of Strategy," Journal of Electronic Imaging, 19 (1), March 2010.
- Reference images: src_imgs.zip (14 MB)
- Distorted images: dst_imgs.zip (365 MB)
- DMOS values: csiq.DMOS.xlsx (0.2 MB)
Reference Images: Thirty reference images were obtained from public-domain sources (mostly from the U.S. National Park Service). The images were chosen to span five categories: Animals, Landscapes, People, Plants, Urban.
Types of Distortions: The distortions used in CSIQ are: JPEG compression, JPEG-2000 compression, global contrast decrements, additive pink Gaussian noise, and Gaussian blurring. In total, there are 866 distorted images.
Protocol: The CSIQ distorted images were subjectively rated base on a linear displacement of the images. Four Sceptre X24WG LCD monitors at resolution of 1920x1200 were calibrated to be as close as possible to the sRGB standard. The monitors were placed side-by-side with equal viewing distance to the subject. The subjects were instructed to keep a fixed viewing distance stable of approximately 70 cm.
All of the distorted versions of each reference image were viewed simultaneously across the monitor array. Each subject horizontally positioned these images across the monitor array such that the horizontal distance between every pair of images reflected the difference in perceived quality between them. As a final step, across-image ratings were performed to obtain a "realignment" of the within-image ratings; this realignment experiment was a separate, but identical, experiment in which observers placed subsets of all the images linearly in space. The ratings were converted to z-scores, realigned, outliers removed, averaged across subjects, and then normalized to span the range [0, 1], where 1 denotes the lowest quality (largest perceived distortion).
Overall the database contains 5000 subjective ratings and are reported in the form of DMOS. Thirty-five total subjects participated in this experiment, but each subject only viewed a subset of the images. The subject pool consisted of both males and females with normal or corrected-to-normal vision. The subjects' ages ranged from 21 to 35.
For more information, please see our paper: E. C. Larson and D. M. Chandler, "Most apparent distortion: full-reference image quality assessment and the role of strategy," Journal of Electronic Imaging, 19 (1), March 2010.