| Recently,many researchers focus on computer vision,especially image semantic segmentation.At the same time,the depth sensors become very popular,and help us collect richer RGB-D images.The appearance of intelligent indoor facilities provide new application for semantic segmentation.However,we need large labeled data for training models.Unfortunately there are little available labeled data,and we have to pay large costs to get them.So our article focuses on RGB-D image semantic segmentation for indoor scenes,and the main works are on:(1)Propose a bottom-up feedback-based weakly supervised segmentation algorithm framework.The framework firstly uses the random forest to get initial labeling,and then provides superpixels by the improved SLIC algorithm.The labeling result is globally optimized by conditional random field based on extracted superpixel features.Finally,use feedback mechanism to solve the problems caused by weak supervision.(2)Explore and optimize the regional feature combinations for random forest.While using random forest to initially label the images,we also explore some different region feature patterns,and use the improved depth of normalization.Finally,the optimal calculation scheme of the feature pattern of random forest is obtained.(3)Present some novel superpixel features for conditional random fields.While using conditional random field,we propose three kinds of features for the irregular characteristics of superpixel,namely,appearance feature,geometric feature and texture gradient feature.Also depth information will be integrated into superpixel segmentation.(4)A feedback-based segmentation mechanism based on image evaluation is proposed.The good result images are selected by the boundary quality and the region quality.And then the selected image will be added back to the training data.By the experiment,we prove our framework can achieve good results in the weakly supervised semantic segmentation,even better than some full supervision algorithms. |
PDF Full Text Request