2022 Machine Learning Lab Public Lecture with Alan Bovik

Refreshments will be provided in the atrium following the talk. Abstract – Every day, hundreds of millions of pictures and videos are captured by inexpert users and streamed and shared on the Internet. Numerous distortions can affect these visual signals: blurs, compression, jitter, shake, noise, judder, over/under-exposure, etc., often combining to create multitudes of composite impairments impossible to model analytically. The problem is made harder because the way that humans perceive distortions depends on the content being viewed: for example, different videos on which identical distortions occur can lie at opposite ends of the perceptual quality scale, because of neurophysiological masking processes. To explain modern methods of measuring perceptual visual quality, I'll explain why video signals are "special," having internal statistical structures that visual systems have optimally evolved to optimally encode and process what we see.