An Approach to Scene Change Detection

Millions of videos are uploaded each day to Youtube and similar platforms. One of the many issues that these services face is the extraction of useful metadata. There are a lot of tasks that arise with the processing of videos. For example, putting an ad is better in the middle of a video, and as an advertiser, one would probably prefer to show the ad in between scene cuts, where it would be less intrusive. Another example is when one would like to watch only through the most interesting or important pieces of video recording. In many cases, it is better to have an automatic scene cut detection approach instead of manually labeling thousands of videos. The scene change detection can help to analyze video-stream automatically: which characters appear in which scenes, how many of them interact and for how long, their relations and importance, and also to track many more issues. The potential solution can rely on different facts: objects appeared, contrast or intensity changed, other colorization appeared, background changed, even sound changed. In this work, we propose the method for effective scene change detection, which is based on thresholding and fade-in/fade-out scene analysis. It uses computer vision and image analysis approaches to identify the scene cuts. The comparison with the PySceneDetect is included. Experimental results demonstrate the effectiveness of the proposed scene change detection approach.