【作者】 王梦颖； 张骏温； 孙波；

【Author】 WANG Meng-ying;ZHANG Jun-wen;SUN Bo;School of Computer and Information Technology,Beijing Jiaotong University;China Railuiay Hohhot Group Co.,Ltd.;

【机构】 北京交通大学计算机与信息技术学院； 中国铁路呼和浩特铁路局集团公司；

【摘要】 语义分割是计算机视觉领域的关键问题之一,越来越多的应用场景需要从影像中得到相关的知识或语义。但是卷积神经网络的不透明性、不可解释和不确定性已经成为制约语义分割实践应用的巨大障碍。文中通过使用Bayesian SegNet网络模型来量化语义分割中的不确定性,以此得到模型对预测结果的不确定性信息,借助不确定性信息来达到辅助决策的目的。Bayesian SegNet网络主要通过进行MC dropout抽样来生成像素级标签的后验分布,研究证明使用Bayesian SegNet量化神经网络的不确定性与贝叶斯神经网络进行概率建模得到的不确定性效果相当。文中在KITTI数据集上对Bayesian Segnet网络模型的不确定性进行了量化,以验证Bayesian SegNet网络结构训练模型得到的不确定性信息来辅助决策的有效性,这一工作对安全关键领域的研究具有重要意义。更多还原

【Abstract】 Nowadays,semantic segmentation is one of the key issues in the field of computer vision,and more and more application scenarios require relevant knowledge or semantics from images.However,the opacity,inexplicability and uncertainty of convolutional neural networks have become huge obstacles restricting the practical application of semantic segmentation.The paper uses the Bayesian SegNet network model to quantify the uncertainty in semantic segmentation,so as to obtain the uncertainty information of the model’s prediction results,and use the uncertainty information to achieve the purpose of assisting decision-making.The Bayesian SegNet network mainly uses MC dropout sampling to generate the posterior distribution of the pixel-level labels.Research has shown that the uncertainty of using Bayesian SegNet to quantify the neural network is equivalent to the uncertainty obtained by the Bayesian neural network for probabilistic modeling.This paper quantifies the uncertainty of the Bayesian SegNet network model on the KITTI data set to verify the effectiveness of the uncertainty information obtained by the Bayesian SegNet network structure training model to assist decision-making.This work is of great significance to research in critical security fields.更多还原