Scribble-Supervised Image Semantic Segmentation Algorithm

Image semantic segmentation is a vital research direction in the field of computer vision,and it is an important part of image semantic understanding.In recent years,the upsurge of deep learning has made image semantic segmentation technology more widely used in the fields of military,meteorology and security.However,image semantic segmentation technology based on deep learning often requires a large amount of labeled data.In order to reduce the workload of data annotation,weakly-supervised image semantic segmentation algorithms have aroused widespread attention,and the weak supervision signals include bounding boxes,points,scribbles and image-level labels.Due to the scarcity of annotation for weakly-supervised image semantic segmentation tasks,the performance of the trained image semantic segmentation model cannot meet the needs of practical applications.In order to solve the problems of sparse and missing supervised labels as well as inaccurate label prediction in weakly-supervised image semantic segmentation tasks,this dissertation has carried out research works for scribble supervised segmentation as follows:(1)Visual pattern correlation propagation for weakly-supervised image semantic segmentation.Aiming at the characteristics of complex regional diversity of the pattern relationship within the image and different-granular feature information contained in the multi-scale feature layers,this algorithm mines the potential spatial/semantic relationships of feature patterns within the same network layer as well as cross different network layers,and propagates the visual patterns through graphical models to form a more robust pattern representation,which further enhances the segmentation performance of segmentation models.(2)Adaptive label diffusion for weakly-supervised image semantic segmentation.In order to solve the problems of scarce supervision information and insufficient segmentation model training in weakly-supervised image semantic segmentation tasks,this algorithm studies the label propagation method based on confident seed regions under the condition of weak annotation,and proposes an improved adaptive label diffusion method which generates the category probability confidence threshold through a dynamic strategy network and forms highconfidence reliable pseudo labels to better guide the update and optimization process of the segmentation network.(3)Progressive inference learning for weakly-supervised image semantic segmentation.With the help of visual pattern correlation propagation and adaptive label diffusion,this algorithm proposes a segmentation framework with progressive inference learning,and conducts information mining on the latent distribution of image self-prior and supervised information,so as to form an alternate update optimization,effectively alleviating the problem of rapid degradation of segmentation model performance as the scope of inference expands.In addition,hierarchical regularization learning is introduced to solve the problem of blurred boundary of the predicted targets,thereby improving the accuracy of the segmentation model.Aiming at the characteristics of weakly-supervised image semantic segmentation,this thesis improves the benchmark method from different perspectives,and the proposed algorithms have been experimentally verified on two public image semantic segmentation datasets and achieved excellent performances,which proves the effectiveness and superiority of the proposed algorithms.