基于视觉显著性的图像特征提取算法  

Image feature extraction algorithm based on visual saliency

作　　者：徐翰文 张闯[1] 陈苏婷[1] XU Hanwen;ZHANG Chuang;CHEN Suting(School of Electronic and Information Engineering,Nanjing University of Information Science&Technology,Nanjing Jiangsu 210044,China)

机构地区：[1]南京信息工程大学电子与信息工程学院,南京210044

出　　处：《计算机应用》2022年第S02期72-78,共7页journal of Computer Applications

基　　金：国家自然科学基金资助项目(61906097);江苏省高校优势学科三期资助项目;江苏省高校品牌专业资助项目。

摘　　要：针对现有特征提取算法中存在运算速度慢、误匹配率高、特征点数量冗杂等问题,提出了一种基于视觉显著性特性的图像特征提取算法。首先,分解彩色图像为HSV矩阵,并改进二维局部熵方法,根据人眼视觉特性,计算出图像的对比度、饱和度、相对亮度和轮廓的信息密度;其次,对归一化后的对比度、饱和度、亮度和轮廓矩阵进行变异系数线性组合,构成图像显著性矩阵;然后,通过非极大值抑制、图像指标均匀化筛选以及剔除小面积区域来进一步分离背景,优化感兴趣区域(ROI);最后,对图像ROI进行分割,再利用改进的尺度不变特征变换(SIFT)算法提取特征。实验结果表明:在显著性区域提取方面,相较于基于定向快速旋转(ORB)和最小凸包的ROI检测算法,所提算法的SP值减小了约15%;在特征提取方面,相较于改进Harris-SIFT算子的快速图像配准算法和信息熵加权的方向梯度直方图(HOG)特征提取算法,所提算法用时分别减少了24%和44%,特征点正确匹配率分别提高了4.8%和3.1%,提取出的特征点数量分别增加了195%和93%。Aiming at the problems that the current feature extraction algorithms have problems such as slow operation speed,high false matching rate,and produced redundant feature points,an image feature extraction method based on visual saliency characteristic was proposed.Firstly,the color image was factorized to HSV(Hue,Saturation,Value)matrix,twodimensional local entropy method was improved and the information densities of contrast,saturation,brightness,and contour of the image were calculated according to the visual characteristics of the human eye.Next,the normalized contrast,saturation,brightness,and contour matrices were used for linear combination of coefficients of variation,and image saliency matrix was formed.Then,the background was separated through non-maximum suppression,image index homogenization screening and removing the small areas,and the Region Of Interest(ROI)was optimized.Finally,the ROI in the image was segmented,and features were extracted by improved Scale-Invariant Feature Transformation(SIFT)algorithm.Experimental results show that in terms of salient region extraction,the SP(Segmentation Processes)value of the proposed algorithm is reduced by about 15%compared with ROI detection algorithm based on ORB(Oriented fast and Rotated Brief)and minimum convex hull.In terms of feature extraction,the time cost of the proposed algorithm is reduced by 24%and 44%respectively,the correct matching rate of feature points is increased by 4.8%and 3.1%respectively,and the number of extracted feature points is increased by 195%and 93%respectively,compared with fast image registration algorithm based on improved Harris-SIFT operator and information entropy weighted HOG(Histogram of Oriented Gradient)feature extraction algorithm.

关 键 词：感兴趣区域 HSV 特征提取 视觉显著性 局部熵 

分 类 号：TP319.4[自动化与计算机技术—计算机软件与理论]