Research On Video Shot Retrieval Based On Visual Salience

Since the1990s, with the development of digital technology and the Internet, theacquisition and dissemination of digital video has become increasingly easy. How to findthe content we need from a broad array of video data has become a hot topic of currentresearch. Content-based video retrieval technology is developed in order to meet thisdemand.The visual psychology studies have shown that human vision has the ability toquickly search for interesting target. In a video shot, only a few targets can cause humanattention. The human visual system is to take advantage of these significant goals todetermine the degree of similarity between different videos. However, the existing videoretrieval techniques commonly use key frame based methods. There are often someinconspicuous targets in a key frame. These targets will inevitably lead to a decline in retrievalaccuracy. This paper attempts to apply visual attention mechanism to video retrieval so that theretrieval accuracy become higher.Firstly, a video shot significant target extraction model is proposed. The model is dividedinto spatial salient regions extraction and temporal salient regions extraction. In spatial salientregions extraction, considering that classic Itti model can not determine the contour of salientregions, a improved spatial salient regions extraction algorithm is proposed. The algorithmanalyses the differences between each pixel with other pixels in color, texture, shape. Spatialsalient map is acquired by overlaying color salient map, texture salient map and shape salient map.The experiments show that this algorithm significantly increase the accuracy of the salient regioncontour extraction. In temporal salient regions extraction, considering that optical flow methodcan not analysis movement information of low texture regional, a new algorithm based on thetrajectory of Harris points is proposed. This algorithm analysis the trajectory of Harris points,extract Harris points belonging to foreground region and determine the contour by Snake model. The experiments show that this algorithm is able to overcome the disadvantage of optical flowmethod and acquire higher accuracy.Secondly, a video shot retrieval experimental platform based on visual saliency wasdesigned and implemented. This platform calculates the similarity of different video shots incolor and shape and find the right result. The experiments show that this algorithm is better thantraditional key frame retrieval algorithm in both precision and recall rate.