PROJECT TITLE :

Detecting and Mapping Video Impairments

ABSTRACT:

Video artefact detection without the benefit of an original reference video is a challenging endeavour. An efficient and innovative dual-path (parallel) excitatory/inhibitory neural network that uses a simple discriminating rule to define a bank of accurate distortion detectors is presented in this paper. Pre-processing each video with a statistical image model makes the learning engine more sensitive to distortion. For visualisation, the complete system is capable of producing full-resolution space-time distortion maps that represent the state of the art in performance. For the first time ever, we have created a full resolution map of artefact detection probabilities using our video impairment mapper (VIDMAP). Eight of the most critical artefact categories found during streaming video source inspection can be accurately detected and mapped by this system's present implementation, including aliasing, video encoding corruptions and quantization, contours/banding and combing. On the whole-image artefact identification challenge, we demonstrate that it is either competitive with or significantly surpasses the previous state of the art. VIDMAP has been trained to detect and map these artefacts and is accessible for public usage and assessment at the following URL: http://live.ece.utexas.edu/research/quality/VIDMAP release.zip.