Weakly Supervised Instance-level Human Segmentation

Image instance segmentation plays an important role in the field of computer vision.With the rapid development of deep learning technology,strong supervised instance segmentation technology based on deep neural networks has made remarkable progress.In order to reduce the dependence of strong supervised instance segmentation technology on high-cost pixel-level manual annotation,research on instance segmentation technology based on weakly supervised learning has gradually attracted researchers' attention.To this end,this paper takes human as an example,adopts a segmentation framework based on deep learning technology,and conducts related research on weakly supervised instance segmentation technology.The specific research work and results are as follows:(1)PRM instance segmentation algorithm using differential evolution.The speed of weakly supervised instance segmentation of PRM method is slow,and the MCG segmentation method that gives object proposals is sensitive to the ranking of segmentation proposals and relies on pixel annotations of other datasets.To this end,this paper uses the Differential Evolution(DE)algorithm for front background segmentation,which eliminates the problem of PRM's dependence on pixel annotation and ranking of MCG proposals,improves the accuracy of instance segmentation,and greatly reduces the time of instance segmentation.(2)IRNET-based instance segmentation with salient feature map generation mechanism.In IRNET method,the class response map can only identify part of the target information.In this paper,a salient feature map(SM)is introduced to highlight the salient regions of the target and an iterative erasion strategy is used to improve the overall information of the target.An improved IRNET instance segmentation algorithm SM?IRNET(IRNET based on Saliency Map)is developed.Experiments show that SM?IRNET can improve the accuracy of weakly supervised human instance segmentation.(3)A semi-supervised instance segmentation algorithm combined with a small number of instance segmentation annotations.Since some instance-level annotations can be obtained for most objects,this paper builds an instance segmentation framework that comprehensively uses existing pixel-level annotations and weakly supervised information.By improving the loss function of the neural network,the network effectively learns the object instance segmentation from multiple information sources.,Which improves the accuracy of instance segmentation.In summary,this paper analyzes the shortcomings of the current advanced algorithm of weakly supervised instance segmentation,and proposes effective solutions to the problems of slow running speed,excessive dependence on pixel labeling,and insufficient information extraction of category response maps.Without a large number of pixel-level instance annotations,the accuracy of instance segmentation of pedestrian targets is improved to a certain extent,the speed of the algorithm is improved,and the foundation for the promotion and application of instance segmentation is laid.