融合目标增强与稀疏重构的显著性检测

发布时间：2018-03-28 12:30

  本文选题：显著检测　切入点：全局颜色对比　出处：《中国图象图形学报》2017年09期

【摘要】：目的为了解决图像显著性检测中存在的边界模糊,检测准确度不够的问题,提出一种基于目标增强引导和稀疏重构的显著检测算法(OESR)。方法基于超像素,首先从前景角度计算超像素的中心加权颜色空间分布图,作为前景显著图;由图像边界的超像素构建背景模板并对模板进行预处理,以优化后的背景模板作为稀疏表示的字典,计算稀疏重构误差,并利用误差传播方式进行重构误差的校正,得到背景差异图;最后,利用快速目标检测方法获取一定数量的建议窗口,由窗口的对象性得分计算目标增强系数,以此来引导两种显著图的融合,得到最终显著检测结果。结果实验在公开数据集上与其他12种流行算法进行比较,所提算法对具有不同背景复杂度的图像能够较准确的检测出显著区域,对显著对象的提取也较为完整,并且在评价指标检测上与其他算法相比,在MSRA10k数据集上平均召回率提高4.1%,在VOC2007数据集上,平均召回率和F检验分别提高18.5%和3.1%。结论本文提出一种新的显著检测方法,分别利用颜色分布与对比度方法构建显著图,并且在显著图融合时采用一种目标增强系数,提高了显著图的准确性。实验结果表明,本文算法能够检测出更符合视觉特性的显著区域,显著区域更加准确,适用于自然图像的显著性目标检测、目标分割或基于显著性分析的图像标注。
[Abstract]:Aim in order to solve the problem of edge blur and poor detection accuracy in image salience detection, a significant detection algorithm based on target enhancement guidance and sparse reconstruction is proposed. The method is based on super-pixel. Firstly, the center weighted color space distribution of the super-pixel is calculated from the perspective of the foreground, and the background template is constructed by the super-pixel of the edge of the image, and the template is preprocessed, and the optimized background template is used as the sparse representation dictionary. The sparse reconstruction error is calculated, and the error propagation is used to correct the reconstruction error, and the background difference map is obtained. Finally, a certain number of suggested windows are obtained by using the fast target detection method. The object enhancement coefficient is calculated from the object score of the window to guide the fusion of the two salient graphs, and the final significant detection results are obtained. Results the experiment is compared with the other 12 popular algorithms on the open data set. The proposed algorithm can detect salient regions accurately for images with different background complexity, and extract salient objects more completely, and compared with other algorithms in evaluation index detection. The average recall rate on the MSRA10k dataset was increased by 4.1%, and the average recall rate and F test on the VOC2007 dataset were increased by 18.5% and 3.1%, respectively. Conclusion A new significant detection method is proposed in this paper, which uses the color distribution and contrast method to construct the salient map, respectively. The accuracy of salient map is improved by using a target enhancement coefficient in the fusion of salient map. The experimental results show that the proposed algorithm can detect the salient region more in accordance with visual characteristics, and the salient region is more accurate. It can be used to detect salient targets, segment objects or annotate images based on salience analysis.
【作者单位】： 辽宁工程技术大学软件学院;
【基金】：国家自然科学基金项目(61172144) 辽宁省教育厅科学技术研究一般基金项目(L2015216)~~
【分类号】：TP391.41
，

本文编号：1676333