| This thesis mainly explores semantic segmentation and UDA issues.Semantic segmentation is an important basic technology in medical image analysis and autonomous driving applications,and UDA can solve the key technology of serious domain offset problems in practical applications of this type of technology.This research first explores the problem of pathological image semantic segmentation in medical image analysis,and proposes a new highperformance semantic segmentation network.On this basis,the focus is on the unsupervised domain adaptation method based on adversarial training in pathological image analysis,aiming to solve the problem of domain shift caused by different staining differences.Finally,this research extends the research object to a wider range of scene semantic segmentation.1)A general process for the recognition and segmentation of pathological image lesions is proposed,and a semantic segmentation network with context awareness and appearance consistency for pathological image lesion segmentation is proposed,which can achieve robust and accurate segmentation.The improved solution achieved the best results in the MICCAI DigestPath 2019 colorectal endoscopic tissue classification and segmentation task.2)A novel confrontation training architecture is proposed.A set of mirrored discriminators are designed for the mainstream segmentation network,so as to ensure the full coverage of the feature map of the backbone network,and ensure the effective transmission of information flow in the network and reduce redundancy.3)A self-training method for instance adaptation of UDA is proposed.A pseudo-label generation strategy adaptively based on instance characteristics is designed.At the same time,for the semantic segmentation task,a region-guided regularization is proposed to effectively use the information of unsupervised data.This scheme has achieved the most advanced performance on the mainstream semantic segmentation domain adaptation benchmark data set. |
PDF Full Text Request