Morphological operators eliminate video noise
Morphological operators eliminate video noise
Until now, linear filtering techniques have been the primary tools used in digital image processing, image analysis, and computer vision. While linear-processing techniques offer satisfactory performance for a variety of applications, many digital image-processing and analysis problems such as non-Gaussian (impulsive) noise filtering cannot be solved using these techniques. They serve only to spread the noise thinly and to blur the image.
To solve this problem, Neal Harvey of the University of Strathclyde (Glasgow, Scotland) is applying nonlinear image-processing techniques to digital video systems. Harvey`s project, funded by the Engineering and Physical Sciences Research Council (Swindon, England) and the BBC Research and Development Department (Kingswood Warren, Surrey, England) uses genetic algorithms to design gray-scale morphological filters and develop a subclass of operators known as rank-order morphological filters.
"Many filters such as median, weighted-order statistic, and stack filters that can filter impulsive noise can be modeled with morphological structures," says Harvey. "With image sequences, you must remove unwanted noise and corruption from the original material," he adds. Therefore, Harvey has devised techniques for the restoration of corrupted 1-in. video-tape-recorder (VTR) sequences that have continuous linear scratches running along the length of the tape. Because such tapes are helically scanned, the VTR head sweeps over the scratch periodically as it rotates and produces a sequence of nonlinear noise or "blips." Currently, the only way to eliminate this effect is to manually paint out the scratches on each frame using a video paintbox. This method is both time-consuming and expensive.
To automate the process, an image is scanned, pixel by pixel, and at each pixel position the values of the pixels immediately above and below are compared to the pixel under consideration. If both these pixels values are greater than a particular threshold value of the pixel under consideration, the pixel is labeled as a potential dark defect. The potential dark-defect binary image is then filtered by a series of close-openings with reconstruction of increasing size, using a one-dimensional, horizontal structuring element.