Convolution Restricted Boltzmann Machine Shape Modeling And Its Image Segmentation Application

Currently there are many image segmentation methods,and these segmentation methods usually rely on the underlying information of the image color,gradient,etc.to segment the image.When the target and background in the image are greatly different or the two have easily distinguishable features,they can often get the correct segmentation result.However,when there are influential factors in the image,such as the target and the background information are similar,the background is complex,the target is obstructed,shadow and so on,rely on the underlying information of the image,often do not get the correct segmentation results.There is a misclassification situation where the target is divided into backgrounds or backgrounds that are divided into goals.While the target shape of the image to be segmented is the most intuitive high-level global feature,it is considered to combine the shape and image segmentation methods to guide the image segmentation process to obtain a correct and complete target.Therefore,the representation of the target shape is crucial.When the shape is multi-class or homogeneous,but there is a big difference in size,pose,etc.,traditional methods such as principal component analysis or active contour model can not be used to express the shape better.To solve this problem,this thesis proposes a new representation method of modeling shape,and studies the application of shape in image segmentation.It uses shape prior helper image segmentation process to improve segmentation results and obtain a complete image target.Based on the above analysis,this thesis studies shape modeling based on deep learning and image segmentation theory,and studies the shape obtained using modeling as a priori information to improve the segmentation effect and improve the accuracy of segmentation.The specific research contents are as follows:(1)Restricted Boltzmann Machines(RBM),Deep Belief Networks(DBN),Deep Boltzmann Machines(DBM),Shape Boltzmann Machines(SBM)are models of deep learning generation.They can model the shape based on the joint probability distribution of the model representation.And they have different structures and different probability distributions.Therefore,the shapes generated by the modeling vary.The best expression is SBM,followed by DBM,followed by DBN,and finally by RBM.However,there is a common problem in this type of model:the image is drawn as a one-dimensional vector and the two-dimensional spatial structure of the image is ignored.Aiming at this problem,this thesis proposes a new probabilistic generation model for modeling deep learning of shapes---convolutional restricted Boltzmann machine.The structure,construction process,mathematical expression and model characteristics of the model are given in this thesis.(2)In order to solve the problem of representation of target shape in image,a shape modeling method based on convolutional restricted Boltzmann machine is proposed.According to the joint probability distribution represented by the model,it can be used to model representations and generate shapes,and the method of model training and generating shapes can be given.This model has the mechanism of weight sharing and local receptive field.It takes the two-dimensional structure of the image as input,retains the spatial structure information of the image,and can extract spatial local information of the image better.Experiments on the Weizmann Horse and Caltech 101 Silhouettes datasets show that the CRBM model has a better ability to model shapes compared to other models of deep learning,and that the model provides a good representation of the details of the modeled shape,and make the shape look more clear and realistic.Using Euclidean distance to measure the difference between the generated shape image and the original shape image(3)In order to solve the problems of level set image segmentation,a shape-based level set image segmentation method is proposed.The target shape is represented by convolutional restricted Boltzmann machine,and establish a shape energy function.Since the level set image segmentation method is based on the energy segmentation method,the shape energy function is combined with the energy function of the original segmentation method to obtain the total energy function containing the shape prior information.The Euler-Lagrange equation corresponding to the total energy function is solved to obtain the minimum value of this energy function,so the shape-constrained image segmentation result is obtained.Experiments on the Weizmann Horse dataset show that the additionof shape information improves the segmentation results and improves the segmentation accuracy compared to the segmentation result without adding shape information.