| Author(s) |
Lin, H., Hosu, V., Fan, C., Zhang, Y., Mu, Y., Hamzaoui, R., Saupe, D. |
| Abstract |
The Satisfied User Ratio (SUR) curve for alossy image compression scheme, e.g., JPEG, gives thedistribution function of the Just Noticeable Difference(JND), the smallest distortion level that can be per-ceived by a subject when a reference image is comparedto a distorted one. A sequence of JNDs can be definedwith a suitable successive choice of reference images.We propose the first deep learning approach to predictSUR curves. We show how to exploit maximum likeli-hood estimation and the Kolmogorov-Smirnov test toselect a suitable parametric model for the distributionfunction. We then use deep feature learning to predictsamples of the SUR curve and apply the method ofleast squares to fit the parametric model to the pre-dicted samples. Our deep learning approach relies on aSiamese Convolutional Neural Networks (CNN), trans-fer learning, and deep feature learning, using pairs con-sisting of a reference image and compressed image fortraining. Experiments on the MCL-JCI dataset showedstate-of-the-art performance. For example, the meanBhattacharyya distances between the predicted andground truth first, second, and third JND distributionswere 0.0810, 0.0702, and 0.0522, respectively, and thecorresponding average absolute differences of the peak signal-to-noise ratio at the median of the distributionswere 0.56, 0.65, and 0.53 dB |