Research On Weakly Supervised Image Segmentation Method Based On Deep Learning

Image segmentation,as the basis of other image processing methods,has always been the hotspot and difficulty in computer vision,which plays an important role in scene analysis,automatic driving and other fields.With the development of artificial intelligent,the deep learning-based image segmentation methods have a qualitative leap in performance and effect compared with the early methods.The existing fully supervised image segmentation methods based on deep convolutional neural network need a large number of pixel-level annotations to improve the segmentation accuracy.So this kind of methods needs a lot of time and economic cost to obtain pixel-level annotations,which restricts the further improvement of segmentation performance and the model generalization ability.To overcome this problem,many scholars try to relax the extent of supervision and propose weakly supervised image segmentation methods based on image-level annotations.The image-level annotations only contain the category information of the image,and have no guidance for the location and contour information of the objects.Therefore,this dissertation starts from the key points and difficulties of this research,focusing on 1)how to use image-level annotations to obtain the location information of the objects;2)how to get the contour/shape priors of the objects;3)how to use image-level annotations to train the segmentation network and other issues.The contents and innovations of each part are as follows.Aiming at the problem that traditional image segmentation methods only use lowlevel features and get unsatisfactory results due to the lack of shape priors,we propose a figure-ground segmentation method based on shape priors.Firstly,we propose using the linear representation of shape and fast directional chamfer matching(FDCM)algorithm to generate more accuracy and class-independent shape priors.This process adopts the data-driven way and omits the model training,which can increase the model generalization ability.Secondly,we introduce shape priors into graph-cuts algorithm,which can improve the segmentation accuracy.The experiments show that our method achieves the better segmentation results on multiple datasets.Aiming at the problem that the early weakly supervised semantic segmentation methods based on deep learning lack location clues,which leads to low segmentation accuracy,we propose a weakly supervised semantic segmentation method based on localizationclues and EM algorithm.Firstly,this method proposes using a classification network to generate localization clues.Secondly,we design a localization clue guiding EM algorithm.This algorithm solves the problem of inaccurate pixel labels estimated in E step and improves the accuracy of the segmentation network.Lastly,we propose a hybrid training strategy.The experiments show that our method solves the problem of no object location priors in image-level annotations,and its segmentation performance is improved compared with the early weakly supervised semantic segmentation methods.Aiming at the problem that the classification network only identifies the most discriminative regions of objects,which results in incomplete pixel labels obtained from image-level annotations,we propose a weakly supervised image semantic segmentation method by combining attention map and saliency map.Firstly,we present a new attention map generating method,and this method can mine most object regions.Secondly,the successive erasing algorithm is designed to detect multiple foreground objects and generate the saliency map.Lastly,we propose a pseudo pixel annotation generating algorithm by combining class-specific attention map with the saliency map,then train the segmentation network.This process not only makes full use of the category information of the attention map,but also adopts the saliency map to supplement the object regions which are not mined by the attention map.The experiments show that our method can get more accurate pseudo pixel annotations,and its segmentation performance is improved.Aiming at the problem that the pseudo pixel annotations generated by combining attention map and saliency map have error marks and can't be corrected in the training process,we propose a weakly supervised image semantic segmentation by an iterative framework combined with Superpixel-CRF refinement.Firstly,a Superpixel-CRF refinement model is designed to introduce contour priors by using superpixels and correct the error marks in the initial pseudo pixel annotations.Secondly,we propose an iterative training framework to progressively improve the segmentation network performance.Lastly,a network alternating training strategy is proposed in the iterative training framework,which avoids the overfitting problem of using a single network and fully integrates the advantages of each network.The experiments show that our method is better than other methods,and it reduces the gap with fully supervised semantic segmentation methods.To sum up,this dissertation studies the problem of weakly supervised image segmentation based on deep learning.Several different kinds of methods for image segmentation are proposed,including a figure-ground segmentation method based on shape priors,aweakly supervised image semantic segmentation method based on localization clues and EM algorithm,a weakly supervised image semantic segmentation method by combining attention map and saliency map,and a weakly supervised semantic segmentation method by combining Superpixel-CRF refinement with an iterative training framework.This work has important theoretical significance and extensive application value for image segmentation task.