基于双向互补学习网络的散焦模糊检测  

Dual direction complementary learning network based defocus blur detection

作　　者：张广强 郑津津[1] 丰穗 苏天成 周洪军[2] Zhang Guangqiang;Zheng Jinjin;Feng Sui;Su Tiancheng;Zhou Hongjun(Dept.of Precision Machinery&Precision Instrumentations,University of Science&Technology of China,Hefei 230026,China;National Synchrotron Radiation Laboratory,Hefei 230029,China)

机构地区：[1]中国科学技术大学精密机械与精密仪器系,合肥230026 [2]国家同步辐射实验室,合肥230029

出　　处：《计算机应用研究》2022年第7期2190-2195,共6页Application Research of Computers

基　　金：国家自然科学基金联合基金资助项目(GG2090090072,U1332130,U1713206);国家自然基金重大仪器专项资助项目(61727809);安徽省重点研究与开发计划资助项目(1704a0902051);国家重点研发资助项目(SQ2019YFC010463)。

摘　　要：散焦模糊检测存在轮廓细节丢失、错分均质清晰区域以及难以处理低对照度渐变区域等诸多问题,针对上述问题,提出一种基于双向互补学习的散焦模糊检测网络,双向学习、逐层融合、互补信息以生成高质量检测结果。双向互补学习网络由特征提取残差模块、双向互补解码子网和融合校正解码子网构成。残差模块提取原始图像的分层级特征;双向互补解码子网同时学习模糊区域和清晰区域的信息,形成互补学习、互补不足;融合校正解码子网则逐层融合成对互补特征图,校正预测误差;此外,所有解码子网均采用分层监督的方式引导网络高效学习。提出的方法在三个公开数据集上F分数分别提升了1.1%、0.1%、1.8%,检测速度达到26.618 fps,超越了现存方法。双向互补学习网络可以有效地挖掘分层级特征和互补标签的信息,快速地生成检测结果。Many challenging problems exist,such as,missing boundary details,misclassifying homogeneous clear areas and the difficulty of dealing with low illuminance regions,in defocus blur detection(DBD).To solve these issues,this paper proposed a dual direction complementary learning network(DDCLNet),via dual direction learning and fusing hierarchical complementary features to obtain high quality results.The network consisted of residual modules,a dual direction complementary decoder subnet(DDCDS)and a fusion correction decoder subnet(FCDS).Residual modules extracted hierarchical features of the source images.DDCDS simultaneously learned clear and blurry information to complement each other.FCDS fused complementary features.In addition,all decoder subnets utilized the supervision mechanism to guide the network to learn efficiently.The F-mea-sure increases by 1.1%,0.1%,1.8%on three public datasets than other methods and the detection speed reaches 26.618 fps,which surpasses exiting methods.DDCLNet can effectively dig out the information of the hierarchical features and complementary labels,to obtain high quality detection results fast.

关 键 词：散焦模糊检测 互补学习 语义特征 结构特征 特征融合 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]