Automatic Colour Enhancement and Scene Change Detection of Digital Video
Jyrki Korpi-Anttila · 2003
Basically digital video is a sequence of still images, displayed at a constant frame rate. Simplest adaptation of still image colour correction algorithm into the digital video is to use the same algorithm frame by frame in the video sequence. However, this kind of approach does not lead to satisfactory results. The sources of the problems are the temporal continuity of the video frames and different gradual transitions, for example editing effects and camera and object movement. Visual defects, which appear in the corrected video are for example flickering, hue changes, brightness changes and edit effect cancellations. These defects are the result of the still image algorithm, which has internally different kinds of thresholds, which adjust correction parameters. There are always small changes in the video sequence. The problems occur when there are small changes in the original video and the changes are still large enough for the correction algorithm to produce little differently corrected image. Even though the changes in the original video are unnoticeable, the changes in the corrected video may be visually noticeable. The temporal continuity of the correction must be made certain and solution is utilizing scene change detection. The object is to detect scene changes and assure that single scene is corrected without any large changes of the correction parameters. The original contrast enhancement and colour balance correction algorithms were based on the RGB-histograms and parameters calculated from the histograms, for example white and black points of the image. It is important that the scene change detection algorithm detects the changes, which results to the eventual changes of the actual correction parameters. In this work, the scene change detection is based on the wavelet transformed histograms. The wavelet coefficients of the histograms of the consecutive frames are compared to calculate difference of the frames. Also separate edit effect detection based on the histogram distribution movement is conducted. Overall result of the scene change detection module is a value between [0, 1], which represents the probability of the scene change. The final correction parameters are the combination of the parameters calculated from the current frame and the parameters used to correct previous frames, i.e. the correction parameters are adaptive recursive filtered before the correction is made. The probability of the scene change is used as an adaptive filter coefficient. If the probability is high then the weight of the parameters calculated from the current frame is significant. If the probability is low then the temporal continuity of the correction is assured by using larger weights for the parameters used in the previous frames. The still image algorithms may be used in the digital video if temporal continuity of the correction is ensured. The filtering of the correction parameters based on the scene change detection is proved to be workable solution. Visual tests of the application presented in this study