SMPTE 2011: Developing a Machine-Learned Video QC Algorithm
It’s darn near impossible for someone to sit at a monitor and spot every video and audio error in a given program. Automated tools fare no better, because they’re difficult to configure, according to
, a master of applied informatics at
in Belgium. However, Vercammen and his colleagues are using machine-learning techniques to develop an algorithm that reliably identifies audio and video errors. He noted that voice-recognition on mobile phones is an example of machine learning.
First off, one must “train the algorithm” he said. The Ghent team used several commercial tools, including an MXF checker and two content verification analyzers—one from Tektronix and another from Interra Systems. They ran several movies through the analyzers and generated data reports. Alerts in the resulting reports were scored on a scale from zero to 100, zero being most critical. The reports were then fed into the algorithm, which in turn learned to identify bad files. It correctly tagged bad files 50 percent of the time using input from just one of the content verification analyzers, but shot up to around 98 percent using input from both.