Research Of Noisy Image Segmentation Based On Level Set And Convolutional Neural Network

As the basis of many image analyses and understanding technologies,image segmentation has been widely and continuously studied by researchers from the entire world.Variational Level Set(VLS)has been a widely used unsupervised learning method due to the advantages of its low computational complexity,reliable mathematical theory with the clear model structure.However,some supervised learning method based on machine learning and deep learning which can realize the extraction and analysis of abstract information such as image semantics,and achieves successful applications in the field of image segmentation.In this thesis,the difficulties of existing VLS methods when processing noisy images and natural images rich in abstract semantics,were overcome,based on the combination of VLS method and machine learning method.The feasibility and effectiveness of the proposed method was verified by multiple experiments using synthetic images and natural image.The details of the work done in this thesis were listed as below:1.A data outlier detection mechanism was constructed based on One-Class SVM,which was introduced into the energy function of VLS method.Thus,the valid information could be obtained by the proposed method under various noise intensities.At the same time,the failure of the existed weight-descend method of strong noise environment was avoided.The evolution process of segment contour could be modeled as the process of minimizing energy function.Finally,the contour stopped near the desired edges.Experiments with a variety of salt&pepper noise environments shows that compared with the traditional noise robust model,the proposed method was more robust to salt&pepper noise without significantly increasing the computational time.Natural images with salt&pepper noise could be segmented.2.Consider the image to be segmented with noise,a noise estimation mechanism was introduced based on convolutional denoising auto-encoder.In this thesis,the existing VLS method based on bias field was combined with the noise estimation mechanism which achieved image noise extraction.Compared with the method of artificially introducing prior statistical information to enhance the robustness of the model,the proposed method has more flexibility and adaptability to images.Finally,a series of contrast experiments on Gaussian noise images prove that the proposed method was more robust to Gaussian noise images and can accurately separate the noise components in the image.3.Based on the unsupervised VLS method,convolutional neural network and deep learning method,a supervised noise robust VLS method was proposed.The proposed method overcame the problem that the traditional method only relies on pixel-level image information,cannot process abstract semantic information and has no learning ability,and improves the performance of the VLS method of the probabilistic noise environment.Experiments show that the proposed method can output practical segmentation results in both synthetic images and natural images in an ultra-high-intensity probability noise environment which is higher than the adaptation of traditional noise robust method.