Weakly-supervised Semantic Segmentation Based On Topic Model

According to semantic segmentation,an important advanced task in computer vision,pixels of image can be assigned into individual labels,providing sufficient materials for image's text description.Notably,text description greatly contributes to the performance of image search engine.Image-level labeling datasets used by weakly-supervised algorithm are much more than pixel-level labeling datasets used by fully-supervised algorithm.Meanwhile,topic model is widely used for text-mining tool for discovery of hidden semantic structures of text dataset.Therefore,this study has been to design a weakly-supervised semantic segmentation based on topic model,indicating an excellent application in enormous image datasets for industry.Firstly,this study described the characters of latent dirichlet allocation(LDA),and put forward weakly-supervised semantic segmentation based on topic model.This algorithm set image labels,salience of image,and Markov random field(MRF)between superpixels as three nodes into LDA model and successfully improve the performance of semantic segmentation.Secondly,this study demonstrated sparse topic model and self-paced frame.The sparse topic model made atonement for the drawback that previous model cannot restrict sparsity of individual image theme explicitly.And self-paced frame can guide to good practice of complex project by obtaining modeling parameters via the practice of simple images.At last,this study focused on the contrast experiment between two open datasets.The results showed that algorithms described previously improved the performance of semantic segmentation significantly,and behaved better than other recent algorithms in the field.