元胞自动机多尺度优化的显著性细微区域检测

发布时间：2018-12-27 20:00

【摘要】：针对由于待检测目标局部区域显著性差异过大造成的细微区域检测失败问题,在贝叶斯理论框架下,提出一种基于元胞自动机多尺度优化的显著性检测方法。首先结合暗通道先验信息和区域对比度在同一张图片的5个超像素尺度空间内分别构建原始显著性图;接着,利用元胞自动机建立动态更新机制,通过影响因子矩阵和置信度矩阵优化每个元胞下一状态的影响力,获得对应5个优化显著性图;最后在基于贝叶斯理论的融合算法框架下得到最终的显著性图。在两个复杂度不同的标准图像数据库上将本文方法与10种主流显著性提取方法进行视觉效果和客观定量数据对比,结果显示,本文算法效果优于现有10种显著性提取方法,其中在公认最具挑战的DUT-OMRON数据库的综合指标F-measure值为0.631 4,平均绝对误差(MAE)为0.132 5,ROC曲线下面积(AUC)为0.892 8,表明本文算法具有较高的准确性和鲁棒性。
[Abstract]:Aiming at the problem of subtle region detection failure caused by the significant difference in local region of the target to be detected, a significant detection method based on multi-scale optimization of cellular automata is proposed under the framework of Bayesian theory. Firstly, combining the prior information of dark channel and regional contrast, the original significant map is constructed in the five super-pixel scale space of the same picture. Then, the dynamic updating mechanism is established by using cellular automata, and the influence of each cell's next state is optimized by the influence factor matrix and confidence matrix, and the corresponding five optimization salience graphs are obtained. Finally, the final significance graph is obtained under the framework of Bayesian theory based fusion algorithm. In two standard image databases with different complexity, the visual effects and objective quantitative data of this method are compared with 10 mainstream salience extraction methods. The results show that the effectiveness of this algorithm is better than that of the existing 10 significant extraction methods. The comprehensive index F-measure value of the most challenging DUT-OMRON database is 0.631 4, the average absolute error (MAE) is 0.132 5g ROC curve area (AUC) is 0.892 8, It shows that the proposed algorithm has high accuracy and robustness.
【作者单位】：北京工业大学信息学部;河南科技学院机电学院;
【基金】：国家自然科学基金资助项目(No.61175087) 北京工业大学2017智能制造领域大科研推进计划“助老智能轮椅床自主测控系统的研究与实现”资助项目
【分类号】：TP301.1;TP391.41

【相似文献】