Robust Semantic Segmentation Model Based On Weakly Labeled Data

Recently,deep learning based semantic segmentation achieves significant success,which heavily rely on large-scale datasets with pixel-level annotation.However,the annotation scarcity hinders the application of deep semantic segmentation in real-world applications.Learning semantic segmentation with limited annotations,such as weakly and semi-supervised semantic segmentation,is a challenging task that has attracted much recent attention.Most leading label-e cient methods employ pseudo labels for self-training.However,the compounding effect caused by inaccurate pseudo-label estimation hurts the generalization and robustness.In addition,the quality of pseudo label is inconsistent among different samples,and thus their importance varies.This paper addresses above issues with in-depth studies in terms of network design,model training and data point reweighting.1)To learn a robust label-e cient model,this paper presents a Self-supervised Low-Rank Network(SLRNet)for labele cient segmentation.The SLRNet uses cross-view self-supervision,that is,it simultaneously predicts several complementary attentive low-rank representations from different augmented versions of an image to learn precise pseudo labels.Specifically,we reformulate the low-rank representation learning as a collective matrix factorization problem and optimize it jointly with the network in an end-to-end manner.The resulting low-rank representation deprecates noisy information while capturing stable semantics across varied views,making it robust to the input variations and reducing overfitting to self-supervision errors.2)To address the inconsistency problem of pseudo label,we propose a dynamic instance indicator(DII)algorithm based on the bi-level optimization.DII considers each weakly-annotated instance individually and learns its weight guided by the gradient direction.In addition,DII is adapted to our dynamic co-regularization framework further to alleviate the erroneous accumulation.Experiments show that our methods can effectively address the error accumulation and improve the generalization and robustness.The proposed methods can be adopted to varied settings of label-efficient semantic segmentation,including weakly,semi-and hybrid supervised tasks,and obtain state-of-the-art results in all these settings on the natural image datasets as well as on the medical image datasets.