计算机工程与应用 ›› 2019, Vol. 55 ›› Issue (3): 179-183.DOI: 10.3778/j.issn.1002-8331.1711-0031

• 图形图像处理 • 上一篇    下一篇

融合边界信息和颜色特征的显著性区域检测

王豪聪,张松龙,彭  力   

  1. 江南大学 物联网工程学院,江苏 无锡 214122
  • 出版日期:2019-02-01 发布日期:2019-01-24

Salient Region Detection Based on Boundary Information and Color Characteristics

WANG Haocong, ZHANG Songlong, PENG Li   

  1. College of Internet of Things Engineering, Jiangnan University, Wuxi, Jiangsu 214122, China
  • Online:2019-02-01 Published:2019-01-24

摘要: 传统的显著性检测方法多利用图像的颜色特征并进行超像素分割作为预处理来进行检测,对于涂抹效应不足、误检测等问题一直没能有效解决。针对涂抹效应不足提出了一种结合图像边界信息及颜色特征的显著性区域检测方法。首先,为了更好地取得图像边缘信息并去除噪声,用多次WMF(加权中值滤波)和简单线性迭代聚类(SLIC)处理源图像,再通过颜色、亮度等信息找出滤波后图像中的自然边界。将得到的边界信息和通过SLIC分割得到的超像素的颜色特征进行融合作为先验概率,以SLIC分割得到超像素位于Graph-based分割得到初步显著图中的概率为条件概率,利用贝叶斯法则得到最终的显著图。在公开数据集MSRA-1000上对算法进行验证,结果表明该算法与7种主流算法相比有更好的查全率和查准率,最高查准率达到98.03%。

关键词: 边界信息, 超像素, 加权中值滤波, 贝叶斯法则

Abstract: The traditional salient region detection method using the color feature of the image and the super-pixels pretreatment has not been able to solve the problem of insufficient smearing effect and mistaken detection. According to the problem, this paper proposes a method for the salient region detection based on boundary information and color characteristics. In order to obtain better image boundary information and remove the noise, the source image is processed by multiple Weighted Median Filtering(WMF) and segmented by a Simple Linear Iterative Clustering(SLIC) to find the natural boundary in the image through the color, brightness and other information. Combining boundary information and the color feature of the super-pixels obtained by the SLIC segmentation, the probability of the super-pixels with the SLIC segmentation locating in the saliency map obtained by Graph-based segmentation is conditional likelihood. Then the final saliency map is gotten in the Bayesian framework. The experimental results which are applied to a public benchmark datasets(MSRA-1000) results show that the proposed algorithm has a better effect on the precision and recall than the seven classical algorithms, the algorithm achieves the highest precision value of 98.03%.

Key words: boundary information, super-pixels, weighted median filtering, Bayesian framework