弱监督学习语义分割方法综述

A Survey of Weak Supervised Learning Semantic Segmentation Method

近年来,深度神经网络的运用使得语义分割技术得到了迅速的发展,然而这些方法在取得更好的分割结果的同时,对训练样本的需求也更多。一般的监督学习语义分割方法要求对训练集中每张图片进行逐像素的标注,这种昂贵的标注需要消耗大量的人力和时间,限制了语义分割算法在工程实际中的应用。弱监督学习语义分割能够使用对象边界框、图片类别标签等标注进行训练,从而大幅度地降低标注成本。文章对弱监督学习语义分割方法进行了分类和对比,并使用电网场景中的图片对方法进行了应用探索,说明了该类方法在工程中实际使用的可行性。

The application of Deep Neural Network(DNN)contributes to the rapid development of semantic segmentation these years.However,to achieve better segmented result,it also raises more requirements for training samples.The normal supervised learning semantic segmentation method requires pixel-wise annotation of each image in training set,which is expensive,time consuming and labor intensive.It limits the application of semantic segmentation algorithm in practice.Weak supervised learning semantic segmentation,which uses bounding box,image category label or other labels for training,could greatly reduce the cost of annotation.In this paper,We classified and compared those methods,and used the images collected from power grid to explore the feasibility of those methods in engineering.