Research Of Image Segmentation Method Based On Few-shot Learning

In the context of deep learning,image segmentation based on convolutional neural network must be supported by a large amount of data.If there are not enough data to train the network,and there are not enough labeled data to support,the network cannot get the segmentation results accurately.Especially in the face of a large amount of data labeling work,not only consumes a lot of manpower but also a waste of time.Another problem is that after the model is trained through the neural network,it is difficult to use the model to predict some new unknown semantic category images in general.In this case,we can only start training the model again.However,with only one or several labeled images,human beings can accurately segment the target in the new unknown semantic image in their mind.The ability of human beings to learn quickly has inspired researchers to study enthusiasm for few-shot image segmentation methods.The purpose of the few-shot segmentation method is to complete the segmentation of new categories of images that have not been seen during the training process with only a small number of labeled images.In this context,this paper has conducted research on few-shot segmentation methods,and provided two ideas for solving few-shot segmentation problems.The main research work of this paper is as follows:(1)This paper proposes a few-shot segmentation method based on an improved guided network.The module is mainly composed of three parts,namely: the guided branch,extracting task representation from the support image and passing them to the segmentation branch,which is used to guide the segmentation of the query image;the segmentation branch,extracting the foreground target of the image under the guidance of the task representation to generate a preliminary rough segmentation map;the fully connected conditional random field,optimizing the segmentation map generated by the segmentation branch and generating a fine segmentation result.(2)This paper proposes another few-shot segmentation method based on the combination of global and local features.When extracting the common information of the support image and the query image,we consider the combination of global similarity and local similarity,and fully consider the interaction between them to solve the problem of few-shot segmentation.The main contributions of this paper include the following two points:(1)In few-shot segmentation methods based on the guided network,a new feature fusion method is proposed.The original support image feature map is globally pooled into a onedimensional feature vector,and then concatenated with query image feature map when the size is restored.The prediction task of dense pixels can be realized only by making some sparse annotations on the support image,that is,randomly pointing several key points in the foreground and background.Finally,the segmentation results are further optimized by introducing the fully connected conditional random field model.(2)In few-shot segmentation methods based on the combination of global and local features,a new attention module is proposed,which includes two parts: a global guidance module and a local guidance module.In the global guidance module,an improved global similarity modeling based on exponential function method is adopted to further enhance the foreground features of the query image.In the local guidance module,the local relationship matrix is considered to measure the local similarity.