Learning To Learn:Weakly-supervised Learning For Image Semantic Segmentation

Image Semantic Segmentation is dedicated to recognizing the content of the image,and identify the category of pixels at each location.The semantic segmentation method based on Fully Convolutional Networks has made good progress.However,this method requires a large number of extremely time-consuming pixel-level annotations.In order to solve this problem,research based on weakly-and semi-supervised has gradually attracted attention.How to make full use of the image Bounding Box annotation,this is a very difficult problem.In the current weakly and semi-supervised algorithms,most of them use handcrafted algorithms to generate image Region Proposals as pseudo labels.This is relatively blunt and rough,and the Bounding Box label information are not fully utilized.In view of the above problems,this paper proposes an image semantic segmentation algorithm based on Learning to Learn for weakly and semi-supervised learning.Based on the fully convolutional segmentation network,Bounding Box is used to annotate infor-mation to learn a general image binary segmentation model,and then use it to generate image Region Proposals.This learning-based algorithm generates image Region Propos-als,which can make better use of the global and local information of the image and the position information of the Bounding Box.In the process of learning the binary segmen-tation model,first proposed a single learning to generate image region suggestions,and then on this basis,combined with the FCN output heat map,proposed multiple iterations of learning,more fully use the image segmentation information.The experimental results on the benchmark dataset Pascal VOC 2012 prove that the performance of the algorithm in this paper has exceeded the current state-of-the-art algorithm.In order to further explore the generalization of the model,this paper migrates the meta-learning model to the task of image Instance Segmentation,and provides a baseline of image Instance Segmentation based on weakly-supervised learning.Under the Mask R-CNN framework,the performance of the weakly-supervised Instance Segmentation algorithm in this paper performs well.