融合背景先验信息及自适应采样的显著性检测  

A saliency detection algorithm via adaptive sampling and background prior information

作　　者：杨琳霞 余映[1] 马玉辉 邓小超 朱信耿 YANG Linxia;YU Ying;MA Yuhui;DENG Xiaochao;ZHU Xingeng(School of Information Science&Engineering,Yunnan University,Kunming 650500,Yunnan,China)

机构地区：[1]云南大学信息学院,云南昆明650500

出　　处：《云南大学学报（自然科学版）》2024年第5期838-851,共14页Journal of Yunnan University(Natural Sciences Edition)

基　　金：国家自然科学基金(62166048);云南省应用基础研究计划(2018FB102).

摘　　要：为了提高计算机视觉系统对视觉场景中显著目标的检测精度,提出了融合背景先验信息及自适应采样视觉显著性检测方法.利用超像素分割提取图像边缘获取场景先验信息,将频域高斯差分特征谱图映射到空间域生成灰度密度散点图,使采样窗口根据散点图自适应移动到前景区域,实现自适应采样.该过程模拟了人的眼动追踪检测显著目标,并采用主成分分析法融合背景先验信息和自适应采样得到的特征,综合提取前景信息获得分辨率更高的显著图.实验结果表明,与其他采样机制的对比,自适应采样方法具有高效性,在ECSSD、HUK-IS、MSRA-5K、SOD公开数据集上与其他13种显著性检测算法对比显示,在各个数据集上MAE平均减少0.01~0.04,F-measure平均提高0.01~0.04,IoU平均提高0.02~0.08,验证了所提算法在显著目标检测的准确性方面具有优势.In order to improve the detection accuracy of computer vision systems for salient objects in visual scenes,an adaptive sampling visual saliency detection method incorporating background prior information is proposed.We use superpixel segmentation to extract image edges to obtain scene prior information.Then,the frequency domain Gaussian difference feature spectrue spectrum is mapped to the spatial domain to generate a gray density scatter plot.Next,the sampling windows adaptively move to the foreground region according to the scatter plot to achieve adaptive sampling.The process simulates human eye tracking to detect salient objects.Finally,We use the principal component analysis method to fuse background prior information and adaptive sampling features to synthetically extract foreground information.The saliency map with higher resolution is obtained.The experimental results show that the adaptive sampling method is efficient in comparison with other sampling mechanisms,and the proposed algorithm is compared with 13 other algorithms on ECSSD,HUK-IS,MSRA-5K,and SOD public datasets.On each dataset,MAE decreased by an average of 0.01−0.04,F-measure increased by an average of 0.01−0.04,and IoU increased by an average of 0.02−0.08,which verify that the proposed algorithm has state-of-the-art in the accuracy of salient object detection.

关 键 词：显著性检测 眼动追踪 自适应采样 背景先验 主成分分析 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]