Research On Image Semantic Segmentation Based On Weak Supervision

Image semantic segmentation is an important task in the field of computer vision and pattern recognition,which is widely used in automatic driving,medicine,smart home and fashion fields.It's goal is to classify all the pixels in the image.However,the traditional full supervised image semantic segmentation learning deep convolution neural network needs a lot of training data marked pixel by pixel,which is limited by the huge cost of marking.Therefore,researchers turn their attention to the weak supervised image semantic segmentation which is easier to obtain marked data.Among them,the commonly used weak labels are bounding box,line,point and image level label.In this paper,the image level weak label with the lowest acquisition cost and the least information is selected to study the image semantic segmentation based on weak supervision.In the existing weak supervised image semantic segmentation algorithms based on image level weak label,most of them are based on the seed growth principle and boundary constraint principle.This paper also takes this as the benchmark,and the specific research contents are as follows:To solve the problem of initial seed sparsity in segmentation algorithm based on seed growth and boundary constraint principle,a weak supervised image semantic segmentation model,GRAD-CAM++-DSRG,is proposed.Based on the DSRG model,this model introduces the GRAD-CAM++ method,which enlarges the scope of the initial seed clue and enhances the supervision information of the segmentation network model.In Pascal VOC 2012 data set,the method is based on a variety of classification networks,and compared with the GRAD-CAM method in the initial seed clue area experiment.The experimental results show that the GRAD-CAM++ method combined with Densenet121 classification network can generate a wider range of initial seed areas.On this data set,the model of GRAD-CAM++-DSRG is compared with the more advanced model of weak supervised image semantic segmentation.The experimental results show that the MIOU value of this model is 58.4%,and its segmentation performance is better than that of the more advanced model of weak supervised image semantic segmentation,which is 0.5% higher than that of DSRG model.To solve the problems of over expansion or insufficient growth of seeds and fuzzy segmentation boundary in the segmentation algorithm based on the principle of seed growth and boundary constraint,a weak supervised image semantic segmentation model Encoder-Decoder-DSRG based on Encoder-Decoder structure is proposed.On the basis of DSRG model,the Encoder-Decoder structure is introduced to improve the information content of low-level spatial features.In Pascal VOC 2012 data set,compared with DSRG model,the experimental results show that the MIOU value of the model reaches 58.4%,which is 0.5% higher than DSRG model,and effectively improves the problems of excessive expansion or insufficient growth of seeds and fuzzy segmentation boundary.The ablation experiments on Pascal VOC 2012 data set and COCO data set are carried out.The experimental results show that the MIOU value of Pascal VOC 2012 data set reaches 59.0%,1.1% higher than DSRG model,and the MIOU value of COCO data set reaches 27.7%,1.7% higher than DSRG model,which effectively improves the segmentation performance.