Dual-dissimilarity measure-based statistical video cut detection

Abstract

Video cut detection is an essential process of temporal continuity-based video applications such as video segmentation, video retargeting, and frame rate up-conversion. The performance of these applications highly depends on the performance of cut detection. This paper proposes an effective and low-complexity approach for detecting video cuts. The proposed method uses two simple dissimilarity measures for video cut detection: inter-frame luminance variation and temporal variation of inter-frame variations over several frames. The first is used to detect abrupt changes, and the second is used to reduce the influence of disturbances, e.g., object or camera motion. The proposed method is comprised of the following three steps. First, it computes the two dissimilarity measures. Then, it combines them using Bayesian estimation and linear regression. Finally, it decides on the possibility of cuts using the combined dissimilarity measure. Experimental results show that the average F1 score of the proposed method was up to 0.252 (37.0%) higher than those of the benchmark methods. Moreover, the algorithmic simplicity of the proposed method reduced the average computation time per pixel by up to 99.8%, when compared with state-of-the-art methods. Thus, the proposed method is superior to existing methods in terms of computational complexity and detection accuracy.