Shot Boundary Detection Using Color Clustering and Attention Mechanism

Shot boundary detection (SBD) is widely used in scene segmentation, semantic analysis, and video retrieval. However, existing SBD algorithms have certain applications processing, but they the following three problems. First, these cannot effectively handle shot boundaries caused by sudden lighting changes. Second, when there are dimly frames video, perform well. Third, object or camera motion also fail to work. To resolve issues, we propose an algorithm with color clustering changes small regions (CCSR) detect transitions, which abrupt gradual transitions (dissolve fade). The main idea behind CCSR compute distance of features preserve spatio-temporal information as much possible. This model has relatively less dependence on threshold parameters sliding windows. Unlike other algorithms, results weaken factors such illumination between adjacent helpful for reducing false detections. Furthermore, utilize attention mechanism improve efficiency accuracy. Finally, evaluated algorithm, was tested a standard TRECVID dataset. experimental suggest that our yields significant improvements precision recall compared current techniques, average improvement 10.35% 8.85%, respectively. Moreover, state-of-the-art prove proposed method improves F-score more than 2.64% computation time over 10%.