特征融合与分割引导的弱监督目标检测

Feature Fusion and Segmentation Guided Weakly Supervised Object Detection

基于卷积神经网络(CNNs)的区域建议生成方法(PRN)是通过实例级注释进行训练所得,也是当前全监督目标检测(FSOD)的重要组成部分。由于实例级注释耗时耗力,而图像级注释相比之下更容易收集,因此仅使用图像级注释的弱监督目标检测(WSOD)引起了众多研究者的关注。当前,WSOD依赖于诸如选择性搜索之类标准的区域建议生成方法,这些方法容易生成大量有噪的建议框,导致其存在无法拟合真实的目标对象。鉴于此,基于卷积特征多层融合以及分割引导策略获取高质量建议框,具体而言,利用卷积网络深层信息进行多层融合,以及边缘信息获取初始的候选建议框,然后通过弱监督语义分割的一致性准则,将分割映射分为水平和垂直两个变量得到目标一致性表示,从而提取高质量的建议框。在PASCAI VOC2007数据集上的实验结果表明,该方法在分类和定位检测中展现了优秀的性能,平均精度(mAP)和定位精度(CorLoc)准确率分别达51.0%、71.2%。

The region proposal generation method(ie,PRN)based on convolutional neural networks(CNNs)is trained through in⁃stance-level annotations,and is also an important part of the current fully supervised target detection(FSOD).Because instance-level annotations are time-consuming and labor-intensive,while image-level annotations are easier to collect,so weakly supervised object detection(WSOD)that only uses image-level annotations has attracted the attention of many researchers.The current WSOD relies on standard region proposal generation methods such as selective search.These methods are prone to generate a large number of noisy pro⁃posal boxes,resulting in their existence that cannot fit the real target object.This paper is based on the multi-layer fusion of convolu⁃tional features and the segmentation guidance strategy to obtain high-quality proposal boxes.Specifically,the deep information of the convolutional network is used for multi-layer fusion,and the edge information is used to obtain the initial candidate proposal boxes,and then through weakly supervised semantic segmentation The consistency criterion divides the segmentation map into two variables,horizontal and vertical,to obtain the target consistency representation,thereby extracting high-quality proposal boxes.The experimen⁃tal results on the PASCAI VOC2007 dataset show that the method in this paper exhibits excellent performance in classification and lo⁃calization detection,with mean of average precision(mAP)and localization(CorLoc)reaching 51.0%and 71.2%accuracy rates,re⁃spectively.