结合轮廓信息递归融合的显著性检测  

RECURSIVE FUSION CONTOUR INFORMATION FOR SALIENCY DETECTION

作　　者：刘岗 林耀进 LIU Gang;LIN Yao-jin(College of Computer Science,Minnan Normal University,Zhangzhou 363000,China;Key Laboratory of Data Science and Intelligence Application,Zhangzhou 363000,China)

机构地区：[1]闽南师范大学计算机学院,福建漳州363000 [2]数据科学与智能应用福建省高校重点实验室,福建漳州363000

出　　处：《南阳理工学院学报》2023年第2期25-31,42,共8页Journal of Nanyang Institute of Technology

基　　金：国家自然科学基金(62076116);福建省自然科学基金(2021J2049)。

摘　　要：针对当前显著性检测算法在检测图像目标边界时,由于轮廓信息不能够及时有效地被利用,导致检测目标效果差的问题,提出一种结合轮廓特征信息的显著性检测算法。该算法以深度残差网络的编解码结构为基础,在自下而上的路径中,首先引用一种融合方式交换轮廓特征与显著特征之间信息的单元模块,然后采用递归结构加强融合进行优化,提升轮廓信息的利用率。最后在此基础上,通过特征提取模块从分阶段的网络模型中提取出最有价值的上层特征,并且与真值图进行监督学习,以生成最优的边界预测。在DUT-OMRON、ECSSD等公开数据集上进行实验,结果表明,相对ITSD、F3Net等算法,该算法能够明显提高检测目标边界的精准度。For the problem that the current saliency detection algorithm detects the boundary of the image object,because the contour information can not be used in time and effectively,which leads to the poor detection effect of objects,a saliency detection algorithm combining the contour feature information is proposed.Based on the decode-encode architecture of deep residual network,the algorithm first designs a unit module that exchanges information between contour features and salient features in a fusion method upon the bottomup pathway,and then uses a recursive structure to strengthen the fusion for optimization and improve the utilization of contour information.Finally,on this basis,the most valuable upperlevel features are extracted from the staged network model through the feature extraction module,and supervised learning is performed with the truth map to generate the optimal boundary prediction.Experiments on public datasets such as DUT-OMRON and ECSSD have shown that compared with algorithms such as ITSD and F3Net,this algorithm can significantly improve the accuracy of detecting object boundaries.

关 键 词：显著性检测 轮廓信息 深度残差网络 监督学习 阶段特征提取 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]