Research On Image Segmentation Based On Finite Mixture Model With Coupled Spatial Informatio

As a key task in digital image processing and computer vision,image segmentation is to classify the pixels in the image and provide a powerful feature basis for the fine understanding and analysis of subsequent image information.Image segmentation has been widely used in industry,remote sensing,medicine,biometrics and other fields,and has increasingly become the focus of scholars.At present,the method based on Finite mixture model has become a popular method in this field because of its good performance in image segmentation.However,the introduction of a large number of model parameters often makes it difficult to select the model and increases the computational complexity.In addition,the current method mainly relies on independent pixels and does not fully consider the spatial neighborhood information of pixels,resulting in discontinuous segmentation boundaries and vulnerable to pseudo similar pixels.In view of the above problems,this paper focuses on the finite mixture model,and carries out in-depth analysis and research from three aspects: target model construction,pixel spatial information exploration,and parameter optimization.The specific work is as follows:(1)Aiming at the problem of weak robustness of Finite mixture model segmentation methods in different scenes,a Gaussian mixture model segmentation method with frequency tuning significance is proposed.The traditional method uses the Gaussian mixture model to extract the significant value of the image in the form of the Gaussian mixture model,which is mainly coupled to the image segmentation value in the form of the significant value of the Gaussian model.By acquiring pixel neighborhood information,the robustness of the model to different scene images is improved.In addition,the proposed method uses the k-means algorithm to select the initial clustering points,and uses the maximum expectation algorithm to solve the model parameters.Experimental results show that the proposed method can not only improve the accuracy of pixel classification,but also realize the robust segmentation of different scene images.(2)Aiming at the problem of low adaptability of model fitting to different image pixel distribution,a bounded generalized Gaussian mixture model with significance spatial information is proposed.The generalized Gaussian distribution is introduced into the finite mixture model as the mixing component to improve the adaptability of pixel fitting.At the same time,the spatial information of the pixel neighborhood is captured by saliency to assist the current pixel to determine its category.It is coupled to the model in the form of weight to solve the problem of edge discontinuity.In addition,considering the pixel boundary support region,a bounded generalized Gaussian mixture model with significant spatial information is formed.And.The model is solved by minimizing the negative log likelihood function by taking the negative logarithm from the maximum likelihood function.Experiments show that the proposed method has good adaptability to the pixel distribution in different ranges,solves the problem of edge discontinuity,and has good denoising performance for noisy images.