Pairwise-Learning Framework for Image Quality Assessment
Tech ID: 29490 / UC Case 2018-613-0
A data-driven, machine-learning approach called the Pairwise-Learning Framework (PLF) that can automatically compute visual error between two images of a given scene in a manner that is consistent with human visual perception.
Creating mathematical models capable of predicting human preferences is a difficult task. Yet, algorithms and various approaches have been attempted. However, these approaches either use hand-coded models which fail to capture the complexity of the human visual system or they use data-driven approaches which are based on small and unreliable datasets. Some pairwise-selection approaches have offered higher levels of success however even these have problems with converting preferences into actual quality scores. Additional problems include noise and other issues that affect estimated scores making them unreliable and difficult to scale to larger datasets. And so, image quality assessments would benefit from an approach that resolves these challenges.
Researchers at the University of California, Santa Barbara have created a data-driven, machine-learning approach called the Pairwise-Learning Framework (PLF) that can automatically compute visual error between two images of a given scene in a manner that is consistent with human visual perception. The output of PLF is the probability that humans would prefer one thing over the other. This approach allows a computer to automatically predict how a human would answer questions such as: Which image appears more clearly? How good is this movie? Would people prefer advertisement A or B? These questions can be answered by simply training the pairwise-learning framework on responses from humans who have been asked to select from a pair of possible choices of the media in question. The unique characteristic of this approach is that despite training on human preference of image pairs, the PLF architecture allows the learned function to predict single image visual error or quality values after the training/optimization process is complete. This approach has proved to be more accurate than the state-of-the-art.
- Significantly increased accuracy
- Consistent with human visual perception
- Video encoding
- Perceptual Image Error Assessment
- Image searching
- Preference predictions