摘要
随着图像采集技术的迅速发展,原始数字图像越来越清晰,已有的协同显著性检测方法在处理这些图像时所需的计算机内存也越来越大,并且伴随着很高的计算复杂性,严重影响了人机交互的实时性。因此,迫切需要一种快速的协同显著性检测方法。提出了一种基于图像分块与稀疏主特征提取的快速协同显著性检测方法(BSFCoS)。该方法在将图像均匀分割成若干个图像块的基础上,从Lab和RGB两种颜色空间上抽取底层特征,再使用截断幂(Truncated Power)的稀疏主成分分析方法进行稀疏主特征提取,以达到在最大程度保留原图像特征的同时减少特征点的数量与属性个数的效果。然后使用K-Means对提取的稀疏主特征进行聚类,并在聚类结果的基础上进行3种基于聚类的显著特征权值的计算。最后,将通过特征融合生成的单幅图像显著图和多幅图像显著图进行组合,以生成协同显著图。在Co-saliency Pairs与CMU Cornell iCoseg两个标准数据集上进行了实验仿真,实验结果表明,与其他协同显著性检测方法相比,BSFCoS在保证检测效果的同时大幅提高了针对多幅图像的协同显著性检测的速度。
With the rapid development of image acquisition technology, the original digital images are increasing and becoming more and more clear. When processing these images, the existing co-saliency detection methods need enormous computer memory along with high computational complexity. These limitations make it hard to satisfy the demand of real-time user interaction. This paper proposed a fast co-saliency detection method based on the image block method and sparse principal feature extraction method. Firstly, the image is averagely divided into several uniform blocks, and the low-level features are extracted from Lab and RGB color spaces. Then truncated power and parse principal components method are proposed to extract sparse principal features, which can remain the characteristics of the original image to the maximum extent and reduce the number of feature points and attributes. Furthermore, K-Means method is adopted to cluster the extracted sparse principal features, and calculate the three salient feature weights. Finally, the saliency map of the single image and that of multi images which are generated by feature fusion are combined to generate co-saliency map. The proposed method was tested and simulated on two benchmark datasets.Co-saliency Pairs and CMU Cornell iCoseg datasets. And the experimental results demonstrate that BSFCoS has better effectiveness and efficiency on multiple images compared with the existing co-saliency methods.
出处
《计算机科学》
CSCD
北大核心
2015年第8期305-309,313,共6页
Computer Science
基金
中央高校基本科研业务费专项资金(NZ2013306)资助
