Weakly-supervised Semantic Segmentation

Semantic segmentation aims at assigning semantic label for each pixel.It has made progress due to development of deep learning and the dense-annotated data.However,time-consuming and laborious annotation cost limits the scale of dataset and categories,preventing general application.Therefore,in recent years,semantic segmentation based on weak supervisions attains increasing attention.Given image-level annotations,existing work mainly utilize Class Activation Maps(CAM)to segment target(called initial seeds)which concentrates on small and sparse discriminative parts,rather than the integral object.To enhance performance,this thesis focuses on seeds region expansion.First,to expand original seeds,the classification network of weakly-supervised semantic segmentation is modified utilizing dilated convolution.The receptive fields of the feature extraction groups are enlarged in order to extend the CAM response.In addition,this paper makes progress like Pooling modification while adopting efficient inference strategies.Furthermore,supplementary data like Saliency is adopted to improve artificial supervision and relevant analyses are elaborated.Through detailed experiments,satisfactory Baseline results are achieved with 60.12% and 61.09% m Io U on PASCAL VOC 2012 validation and test set,respectively.Based on the contrast,this paper chooses to focus on the fundamental and vital essence of weakly-supervised semantic segmentation,which is the performance of initial seeds generation.Second,due to the seeds simply focus on discriminative regions,it is a challenge to spread seeds to the integral object.To tackle this problem,we propose a Cascade Semantic Erasing Network(CSENet)to expand seeds effectively and reasonably.In particular,CSENet sequentially stacks the semantic erasing stage to erase discriminative areas progressively.It forces the network to discover relevant feature responses for non-discriminative object districts.Moreover,CSENet directly erases seeds on the CAM,which have stronger semantics,rather than on the Intermediate Feature Maps(IFM).With semantic guidance,this erasing strategy correctly spreads seeds regions to the intra-class regions and meanwhile,prohibits from extending to the unexpected inter-class areas.Extensive experiments demonstrate the effectiveness of proposed CSENet on expansion efficiency with 62.31% and 63.37% m Io U on VOC validation and test set,respectively.Finally,spatial erasing methods regard the channel-wise representations as an integration for a certain pixel,losing all the channel features and leading to a semantic confusion with two typical problems: they lose various channel patterns for identical category and amplify the initial false patterns.Considering diverse features encoded in channels,the Adversarial Channel Dropout(AC-Dropout)is proposed,which provides the channel instead of spatial perspective for region expansion.Specifically,channel predictions are separated and the method drops high response channels to impel model to activate the remainder channels,exploiting regions by pattern diversity.In addition,this adversarial manner inspires the competition among channels,thus correcting the original false alignments.Experiments demonstrate the effectiveness of proposed AC-Dropout compared with spatial erasing.The method performs 62.76% and 63.54% m Io U on VOC validation and test set,respectively.This thesis mainly focuses on the research of initial seeds generation from classification model for weakly-supervised semantic segmentation.By enhancing the artificial supervision,promising results are achieved compared with existing methods.Proposed approaches with corresponding intuition provide the brand-new perspective for weakly-supervised semantic segmentation,which is meaningful in the research field.