Scene-Change Detection using Locality Preserving Projections

Scene-Change Detection / Video Shot Boundary Detection is an important pre-processing step in applications like automatic content analysis or auto-captioning etc. The goal of scene-change detection is to divide the video stream into a set of meaningful and manageable segments (shots) that can be used in subsequent processing. Camera action and special cinematic effects pose a great challenge in scene-change detection. This article proposes a robust scene-change detection method based on Dimensionality Reduction Techniques. The main contribution of this article is a scene-change detection method that handles gradual changes efficiently. The proposed method exploits the similarity preserving property of Locality Preserving Projection (LPP) to detect the dissimilar frames in the video stream. The performance of the LPP based proposed method is compared with the Principal Component Analysis (PCA) based method. The experiments done on movie scenes, television shows and animated videos show that the proposed method outperforms PCA based methods when video has challenging scene-changes like gradual dissolving or fade-in, fade-out.