The Research Of Weakly Supervised Semantic Segmentation Algorithm For Image Based On Seed Growth And Constraint

Artificial Intelligence is getting closer to people's daily life.Semantic segmentation is a fundamental problem in the field of computer vision,and it has been widely used in automatic driving,robot's visual perception,clothing classification and geological analysis and so on.Most of the traditional semantic segmentation methods are full-supervised,and the cost of obtaining pixel-wise labeled training data is very high.Therefore,some researchers have turned their eyes to the research of weakly supervised semantic segmentation algorithm which is considered to have more potential application value.Following the principles of seed growth and constraint is a popular way in existing algorithms for weakly supervised semantic segmentation.We focus our work in the existing method named SEC(seed expand and constraint)and improved its performance with following optimization methods.(1)A model named MPCSEC(multiple priori constraint SEC)is proposed,in which multiple priori constraints including suppression,foreground and background cues are introduced to guide the growth of the seeds,it aims at helping overcome the over-growth or under-growth problems in SEC and make the seeds have more reasonable expanding ability.(2)An image preprocessing layer is added,in which we use Adaptive Contrast Enhancement for image and Image Affine Transformation as preprocess ways to improve the quality and quantity of training data so that the model can focus much more attention on discriminating features of target objects.And with richer seed cues,the reasoning ability of the model can be enhanced as well,and it can also make up the shortcomings of sparse and intermittent in it's initial extraction of seed pixels of the SEC method.(3)Optimization methods mentioned at(1)and(2)are integrated to form a stronger MSOSEC(multiple stage optimization SEC)model which perform better than SEC.(4)The MSOSEC model is proposed to be pruned and fine-tuned so that we can obtain a more streamlined and efficient model to satisfy the need of weakly supervised semantic segmentation in mobile and real-time applications for the coming future.At the same time,the effectiveness of the proposed methods is verified by experiments.The experimental results have shown that with the final proposed MSOSEC model,the segmentation accuracy improved by 3.8% compared to the existing SEC.And after pruning,the accuracy loss of the MSOSEC model is 0.3%,and at the same time,the scale of model parameters reduced by 16.5% and the time for segmenting a single image in average is decreased by 25%.We obtain better performance for semantic segmentation in the situationwith only image-level labels.