Texture-based two-stage shot boundary detection in videos

Abstract

In recent years, shot boundary detection (SBD) has become an essential component of video processing, enabling applications such as video indexing, summarization, and content retrieval. However, the task remains challenging due to frequent false positive detections caused by illumination variations, motion changes, and diverse editing effects. To address these challenges, this paper presents a novel two-stage SBD framework that leverages local quad pattern (LQP) histogram features for precise transition detection. In the first stage, histogram feature vectors are derived by counting the occurrences of LQP codes (−1, +1, 1, 0), and abrupt transitions are identified using the Euclidean distance between consecutive frames. In the second stage, mean values of each histogram bin are computed for consecutive frames, and a similar distance-based approach is applied to refine detection accuracy. A transition frame is confirmed as a shot boundary only if both stages detect it, thereby reducing false positives. The proposed method is evaluated on the TRECVid 2001 and 2007 benchmark datasets, and experimental results demonstrate its superior performance compared to existing algorithms.