Image Quality Estimation

Images and videos are ubiquitous today. The share of bits representing visual signals is huge and ever growing. Storage and transmission of these signals require bit rate reductions through compression. However, with decreasing bit rate, distortions that are visible for the human eye are introduced into the signal. In order to automatically evaluate and optimize the performance of compression and transmission systems, a metric for image or video quality is necessary. As humans are typically the ultimate receiver of visual signals, it is crucial for such a metric to relate to human visual perception and to predict the visual distortion perceived by humans reliably.

In our research we investigate and develop computational models to reliably predict visual quality as perceived by humans. Methodologically, we work with traditional signal processing techniques, but we also pioneered the usage modern deep learning approaches in the visual quality domain. Our past successes for instance include the development of a convolutional neural network that jointly learns the relative importance of local and global visual image quality or the conceptual identification, introduction and mathematical modeling of distortion sensitivity which bridges from psychophysically motivated human visual sensitivity to natural image statistics.

Our approaches were shown to be highly feasible to optimize visual computing systems. When incorporated into modern video compression schemes such as VVC for bit allocation, our methods and models can reduce the bit rate by 16% over conventional bit allocation techniques.