A Robust Clustering Algorithm for Video Shots Using Haar Wavelet Transformation

Automatic clustering of video shots is an important issue of video abstraction, browsing and retrieval. Most of the existing shot clustering algorithms need some prior domain knowledge or thresholds to obtain good clustering results, and they also have to face the difficult task of choosing proper initial cluster centers. To resolve the discommodious problems for users, this article proposes a robust unsupervised shot clustering algorithm which is called CAVS (Clustering Algorithm for Video Shots). In CAVS, multiresolution analysis and Haar wavelet transformations are first applied as a dimensionality reduction approach for the high-dimensional feature vectors of shots. Then CAVS performs on the remained subspace and merges the most similar shots into one cluster by the iterative merging procedures. The iterative merging procedures are repeated until a novel stop criterion based on the theory of Fisher Discriminant Function is satisfied, and the clustering results and the number of clusters are obtained without any parameters.