Research On Divide-and-Conquer Strategy Based Weakly Supervised Image Semantic Segmentation

With the development of computer vision technology,image semantic segmentation,as one of the core issues,has been paid more and more attention by academic and industry.The purpose of image semantic segmentation is to divide each pixel in the image into pre-defined semantic categories.Through its study,on the one hand,it can help to understand the human visual mechanism to assist in the exploration of image understanding and scene perception and other higher-level visual tasks,on the other hand,it can also provide theoretical and technical support for a wide range of practical applications such as autonomous driving and image search.The existing image semantic segmentation algorithms based on image-level tags restore the missing label information as a whole,and typically generates pixel-level pseudo-labels data for the image first,and then uses these pseudo-labels data to train the algorithm model.Although these methods discard the need for real pixel-level tags data,greatly reduce the cost of model training.But its recovery of label information is not complete,so that the generated pseudo-labels always exist mislabeled areas,and the edges between objects are not clear enough,resulting in bad segmentation result.In view of the above problems,this paper puts forward a divide-and-conquer based method to recover the missing label information,builds the dual-feedback network model,improves the quality of the existing pseudo-label data,corrects the mislabeled area in the pseudo-label data and refines the edge of the split between objects.Compared with the existing methods,this paper observes the missing label information from a new perspective and further divides it into low-level physical location information and high-level semantic location information,and recovers independently.Under the guidance of this idea,this paper proposes that the super-pixels based pseudo-labels walking mechanism and the pseudo-labels updating mechanism recover two different location information respectively,and construct a dual feedback loop network model by transforming the two mechanism into the form of network feedback loop.The experimental results show that the segmentation performance of the dual-feedback network model in this paper is better than that of the existing method on the two publicly segmentation data sets.In order to further verify the effectiveness of the idea above,thispaper also verifies the working principle of two different mechanisms by qualitative and quantitative evaluations,and shows the improvement of pseudo-labels data.