概率框架下多特征显著性检测算法

显著性检测是计算机视觉的一项基础问题，广泛地用于注视点预测、目标检测、场景分类等视觉任务当中.为提升多特征条件下图像的显著性检测精度，以显著图的联合概率分布为基础，结合先验知识，设计一种概率框架下的多特征显著性检测算法.首先分析了单一特征显著性检测的潜在缺陷，继而推导出多特征下显著图的联合概率分布；然后根据显著图的稀有性，稀疏性，紧凑性与中心先验推导出显著图的先验分布，并使用正态分布假设简化了显著图的条件分布；随后根据显著图的联合概率分布得到其极大后验估计，并基于多阈值假设构建了分布参数的有监督学习模型.数据集实验表明：相比于精度最高的单一特征显著性检测方法，多特征算法在有监督和启发式方法下的平均误差降低了6.98%和6.81%，平均F-measure提高了1.19%和1.16%；单幅图像的多特征融合耗时仅为11.8ms.算法精度较高，实时性好，且可根据不同任务选择所需的特征类别与先验信息，能够满足多特征显著性检测的性能要求.

Saliency Detection with Multi-features in Probability Framework

Saliency detection is a fundamental issue in computer vision.It is widely applied in fixation prediction,object detection,scene classification,and other visual tasks.In order to improve the precision of visual saliency detection with multi-features,a multi-feature integration algorithm is proposed based on the joint probability distribution of saliency map and combined with priori knowledge.Firstly,the potential defects of single feature saliency detection are analyzed,and the joint probability distribution of saliency maps with multiple features is deduced.Secondly,the priori distribution of the saliency map is deduced based on the rarity,sparsity,compactness and center priori of the saliency map,and the condition distribution of the saliency map is simplified based on the assumption of normal distribution.Then the maximum a posteriori estimation is obtained from the joint probability distribution of the saliency map,and a supervised learning model of the distribution parameters is constructed based on the multi-threshold hypothesis.Experiments show that compared to the highest-precision saliency detection method on single feature,the mean average error of the multi-feature algorithm under the supervised and heuristic method is decreased by 6.98% and 6.81%,and the average F-measure is improved by 1.19% and 1.16%.And the multi-feature integration of single image takes only 11.8ms.The algorithm has high accuracy and real-time performance,and can be combined with the required features and different prior information according to the task.It meets the requirements of saliency detection with multi-features.