Abstract: This paper proposes a novel video saliency detection method. Firstly, in order to extract features with high confidence in the video sequence, a simple frame selection criterion is proposed according to the initial saliency map of the input frame and the input frame, and the simple selection criterion is used to select the video sequence to easily and accurately extract the foreground object, it gets robust foreground background label from simple frames. Then the image is sub-pixel segmented, and the time-space feature and foreground tag input integrated learning model is extracted. After multi-kernel SVM integration learning, a pixel-level saliency map is finally generated, and the motion feature is spread to the entire video set. Experimental results of various video sequences show that the algorithm is superior to the traditional significant detection algorithm in qualitative and quantitative.

Key words: simple frame selection, saliency detection, multi-kernel SVM bootstrap learning

摘要： 提出了一种新颖的视频显著性检测方法。为了提取视频序列中具有高置信度的特征，根据输入帧和输入帧的初始显著图提出一种简单帧选择标准，并使用该简单选择标准挑选出视频序列中比较容易且准确提取前景对象的帧，从简单帧中获得鲁棒的前景背景标签；将图像进行超像素分割，提取时空特征与前景标签输入集成学习模型，经过多核SVM集成学习，最终生成像素级别的显著图，并且由运动特征扩散到整个视频集。各种视频序列的实验结果表明，该算法在定性和定量上优于传统的显着性检测算法。

关键词: 简单帧选择, 显著性检测, 多核SVM集成学习