Research On Cosegmentation Based Weakly Supervised Semantic Segmentation Methods

Semantic segmentation is the fundamental research in computer vision,and is the basis of many high-level semantic analysis tasks such as visual understanding and behavior recognition.Semantic segmentation is a difficult task due to the variability of foreground and the interference of complex background.Traditional methods based on artificial designed feature and statistical machine learning can not effectively extract semantic object regions.Recently,deep learning has provided a new way for image semantic segmentation.With a large number of segmentation annotations,deep learning based semantic segmentation methods can effectively achieve image semantic segmentation.However,in practice,there is often no sufficient pixel-level training data,so researchers turn to weakly supervised semantic segmentation,which achieves semantic segmentation by extracting object prior from image-level annotations.Because of the simplicity of annotation and the easy acquisition of large image data,weakly supervised semantic segmentation has the advantages of low cost and abundant training data.Therefore,the research on weakly supervised semantic segmentation is significant for breaking through the bottleneck that the current fully supervised semantic segmentation methods rely much on training data.However,there is a huge gap between accurate pixel-level segmentation information and rough image-level information provided by image-level annotation,which makes the image-level annotation based semantic segmentation a difficult task.Specifically,the difficulties are as follows:(1)The segmentation model can not learn accurate object prior from the rough object prior provided by image-level labels.(2)Due to the diversity of foreground and the complexity of background,it is a common problem for weakly supervised semantic segmentation methods to segment object regions from rough object priors.(3)Due to the lack of prior in single class labels,how to combine multiple class labels to achieve weakly supervised semantic segmentation is a difficult task.In view of the above problems,this paper carried out a series of studies as follows:1.In order to solve the problem that the object prior in image-level labels is too rough to accurately learn the segmentation model,a weakly supervised semantic segmentation method based on objectness prediction and active learning is proposed.By constructing a cosegmentation model based on the shortest path search of directed graph,the transformation from image-level labels to pixel-level pseudo labels is achieved,while a few number of human interaction is involved.With a few human interaction,the prior extraction of object and the generation of segmentation training data are effectively completed.2.In order to solve the common problem of object segmentation based on foreground prior in weakly supervised semantic segmentation,a foreground probability map based segmentation model is proposed,and an optimization strategy based on iterative segmentation is proposed.The accurate object region segmentation based on rough object prior is achieved,and the accuracy of foreground segmentation is improved.3.In order to solve the problem of insufficient prior information in single group image cosegmentation,this paper proposed a multi-group image cosegmentation framework for weakly supervised semantic segmentation based on foreground segmentation network.A fusion method based on foreground region and discriminative probability map.is proposed.By optimizing the energy function of multi-group image cosegmentation,the performance of weakly supervised semantic segmentation is effectively improved.