Research On Image Instance Segmentation Based On Weakly Supervised Learning

Instance segmentation,which could implement classification,localization and segmentation for each object of interest,has been widely applied in the field of video surveillance,autonomous vehicles and medical imaging.Compared with conventional instance segmentation,instance segmentation based on weakly supervised learning could greatly decrease financial and time cost.It has been improved greatly in recent years.However,problems like the following still exist: On the one hand,existing methods ignored the effect of the quality of generated pixel-level labels.This caused some invalid masks were used to train networks.On the other hand,existing methods were not improved to adapt weakly supervised labels.To address aforementioned problems,this thesis proposed a complementary two-channel instance segmentation network based on weakly supervised learning.In this architecture,there is a large object branch which is used to train detection and segmentation networks simultaneously with valid generated labels.In the process,detection task and segmentation task promote each other,and invalid labels are incapable of interfering with network learning.In addition,a small object branch is added to handle the small objects without valid labels.It firstly implements detection for small objects,and then uses Grab Cut to segment these detection results.This branch could spare small objects from the supervision of the pixel-level labels.Finally,the predictions of two branches are fused to make full use of two branches' complementary.Experimental results indicate that the proposed method outperforms existing state-of-the-art methods,and achieves significantly performance improvement both on the small objects and large ones.Weakly supervised instance segmentation network was improved by above method for the first time based on the quality of generated labels,and achieved prominent performance.However,its network structure was relatively complicated.Besides,it also cannot take full advantage of the limited supervised information.Therefore,a novel weakly supervised instance segmentation method based on two-stage transfer learning is proposed to address the problems mentioned above.In the first stage,a network-based transfer learning strategy,which could transfer the parameters from object detection network to instance segmentation network,is used to effectively utilize all useful information.For further improving the instance segmentation performance of small objects,a feature-mapping-based transfer learning strategy is proposed in the second stage to transfer the detection feature to segmentation domain.Experimental results demonstrate the effectiveness of two-stage transfer learning.Compared with state-of-the-art methods,the proposed method has better performance and more lightweight network structure.