基于类别概率反向传播机制的弱监督语义分割  

Weakly supervised semantic segmentation based on category probability back propagation mechanism

作　　者：李良御 Li Liangyu(College of Physics and Information Engineering,Fuzhou University,Fuzhou 350116)

机构地区：[1]福州大学物理与信息工程学院,福州350116

出　　处：《电气技术》2020年第4期80-84,共5页Electrical Engineering

摘　　要：本文利用带语义信息的边界框作为弱监督标注,借助目标边界框作为先验线索,定位分类网络中属于目标但激活值较弱的特征点。通过概率反向传播机制的方式寻找各卷积层之间神经元节点的关联性,获得目标较完整的类别注意力图。另外,结合图像超像素算法,通过填充率选择策略改善注意力图在边缘处的粗糙分割效果,生成最佳的类别掩膜。本文提出的方法改善了以往注意力机制的定位不完整,并在PASCAL VOC2012分割数据集上获得了64.8%的mIoU分值结果。The bounding box with semantic information is used as the weak supervised annotations,and the object bounding box is used as a priori clue to find the feature points that belong to the target object but have weak activation value in the classification network.The correlation of the neuron nodes between the convolution layers is found by probability back propagation mechanism,and a complete class attention map of the object is obtained.In addition,by combining the image super-pixel algorithm,the rough dividing effect at the edge of the attention map is improved by the filling rate selection strategy,and the optimal category mask is generated.The extensive experiment results show that the method proposed method improves the integrity of the positioning of attention mechanism,and obtains 64.8%mIoU score results on the Pascal VOC2012 segmentation dataset.

关 键 词：反向传播 弱监督 语义分割 注意力机制 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP18[自动化与计算机技术—计算机科学与技术]