Research On Shape Model Based On Deep Learning And Image Segmentation Application

In order to better understand and analyze the object in the image,it is necessary to separate it from the complex image content.That is image segmentation problem.Image segmentation is an important research content in the field of pattern recognition and computer vision.It is a key step from image processing to image understanding and image analyzing.It has been widely used in medical diagnosis,video monitoring,remote sensing image analysis,image object detection and extraction,and so on.At present,there are many methods to solve the problem of image segmentation,but so far there is still no universal method.In practical applications,the segmentation results are easily influenced by complex background,occlusion and other factors,and it is an effective way to use shape prior to assist the segmentation of object in the image.Based on the image segmentation algorithm and deep learning theory,this thesis gives a study of the shape model based on deep learning.Further,this thesis focused on the problem of how to exploit shape priors for image segmentation problem.The major work and innovations lies in the paper as follows:(1)For shape modeling,a shape representation model based on deep learning is proposed,which is a generative model to represent the shape and generate shape.According the structural characteristics of the model,the construction method and the training method of the model are given.A method of shape prior representation and estimation is established by using the underlying features and multi-layer high-level features of the shape.The experimental results show that the model can not only fit the training shape well,but also generate the shape different from the training samples.When training shapes are multi-category,the model can also define the multimodal distribution.(2)Based on traditional image segmentation method using underlying data features such as image spectrum information,combined with the deep learning shape model,an image segmentation model based on conditional deep Boltzmann machine is proposed.The model segment image by using the image underlying data features and shape priors information.The deep Boltzmann machine in this model is used to extract the distribution structure of the output space data.Therefore it can provide the shape prior information for image segmentation,and then improve the accuracy of image segmentation.The experimental results show that the proposed model takes advantage of the shape prior information and considers the high-order correlation of the output space.Therefore,a good result is obtained when the natural image with the interference factors such as occlusion and background chaos is segmented.(3)A Markov random field based image segmentation associate with deep learning shape prior is proposed.The method uses the deep learning shape model to establish the representation and estimation of the object shape prior,and uses symbolic distance function in level set to measure the similarity of shape and define the shape energy term.It is combined with the original energy function to obtain the segmentation energy function including appearance information and shape prior information.And then the segmentation energy function is minimized by graph cuts to get image segmentation result.Through the combination of shape prior information and appearance information,the effect of superposition enhancement is produced and the influence of the interference factors on the segmentation result is reduced by using the shape prior constraint.The experimental results show that the introduction of deep learning shape prior information in the Markov random field image segmentation using the image underlying data information can effectively improve the segmentation result and get the correct object.