Published in Annual Meeting, Vision Sciences Society, vol.56.409 May 2014.
Quantifying the visibility of distortions in photographic images is of fundamental importance in many fields. The most commonly used error measure, root-mean-squared error (RMSE), is a poor predictor of perceived image quality, because the visibility of a given distortion is critically dependent on its spatial structure and its relationship to the underlying image. Specifically, some patterns of distortion affect perceived image quality less because they are masked by the image content. Wang and Simoncelli (IEEE Proc. ICIP, 2005) proposed a metric based on a weighted sum of squared differences in which a family of image-dependent distortion components receives smaller weights than other components. These image-dependent components are derived from the image content as linear approximations to naturally-occurring distortions, such as spatial translations and changes in mean luminance or contrast. Here, we introduce a psychophysical method to estimate the appropriate weights for these image-dependent distortion components. Stimuli consisted of natural images, and distorted versions generated for each image-dependent component, D (e.g., horizontal translation), each possible weight, wD, for that distortion, and one of a fixed set of RMSE values. Distorted images were generated that maximized image quality according to the metric (using wD) for each fixed RMSE value. Observers were required to detect the distorted version in a 3-interval forced-choice task (two copies of the original and one distorted image). Detection thresholds were measured as a function of RMSE for each D and wD. We predicted, and indeed found, that RMSE detection thresholds obey an inverse-U-shaped function of wD. The peak of this curve occurs at a value of wD for which observers are least sensitive to these quality-metric-minimizing distortions, providing an optimal value for wD. Thus, our technique may be used to tune the image quality metric to reflect human sensitivity to image distortion.