Video stills the criminal star

February 24, 1995

Aisling Irwin reports from the annual meeting of the American Association for the Advancement of Science in Atlanta. The video footage of a murderer's getaway car, which was so blurred that not even the car's make was clear, was so improved by a new scientific technique that police discovered it was one of only ten possible cars in America.

The computational technique - based on the principle of least total variation - has also led to the exoneration of a man on trial for murder when it proved that what was thought to be a large "afro" hairstyle on a video was in fact a base-ball cap.

Leonid Rudin, consultant to the Los Angeles police department and visiting professor at the University of Los Angeles, devised the technique in order to make use of the masses of video film that could help solve crimes but is of poor quality. This could be because the crime happened out of the range of the camera's focus, the tape was old or there was dirt on the lens.

Efforts to process such images have always led to blurring and other artefacts. This is because they try to smooth out the whole picture. But the smoothing also removes the crucial contours that mark out the features of the face.

The new technique takes a video still and plots it as a three-dimensional bar chart. The chart still looks like the face on the video, with high bars where the image is white and lower bars in places where the face grows darker. A poor image will have lots of alternating high and low bars (oscillations). A clear image will be smoother but will have sudden discontinuities - for example where the edge of the face suddenly becomes the background shadow.

The key to the technique is that it can discern the discontinuities that are essential to the image.

The image is modelled mathematically. It is found to obey "natural" non-linear principles, such as the principle of least total variation. The problem of restoring the image can be solved numerically, using non-linear partial differential equations. The computer searches for a solution to the mathematical function that is smooth with the least number of oscillations.

The technique has now been accepted in the courts.