Research On Coding Of Elemental Image Array Based On Gaussian Mixture Model

Integral imaging can reproduce a large-scale of the real 3D scene with continuous parallax and real color.However,it is difficult to store and transmit elemental image arrays which can reproduce 3D scene in integrated imaging,which limits the development of related applications.In order to achieve an efficient coding of elemental image arrays,we try to fit the pixel information of the image using Gaussian mixture model(GMM).The similarity among Gaussian distributions can be analyzed to encode elemental image arrays.As the basis of the coding of elemental image arrays based on GMM,an image coding algorithm based on GMM is proposed in this paper first.Then,GMM is introduced into the coding of elemental image arrays.The research breaks the idea of the traditional image coding which achieves compression by removing the correlation of an image based on pixel blocks.In order to utilize the great correlation within the elemental image array,the research represents elemental-images using Gaussian mixture model.An efficient coding of elemental image arrays,which is based on Gaussian mixture model,is proposed in this paper.At present,there are few studies on image compression based on Gaussian mixture model.And there are still some problems for these model-based image coding algorithms.First,it is difficult to determine the number of distributions needed during the prediction.Second,the predicted image is not clear enough using the common Gaussian mixture model.Third,the correlation among the elemental-images in the elemental image array is not utilized.The correlation among the parameters of the distributions is not analyzed.Fourth,the coding performance still needs to be improved.In view of the shortcomings of the existing researches,the determination of the number of distributions are addressed in this paper.The quality of the predicted image is also improved.And the covariance parameters are analyzed and match using the proposed method in this paper.The main contributions and innovative points of this dissertation mainly contain the following five aspects:1.Considering the uncertain number of distributions in Gaussian mixture model,we propose GMM Model Optimization(GMO),which is suitable for image coding algorithm.The method considers both the quality of the predicted image and the encoding cost.By calculating GMO values with different numbers,the number of distributions corresponding to the largest GMO is selected as the optimal number of distributions to ensure the uniqueness of it.In this method,the optimal number of distributions for image coding is determined by quantitative calculation,which can avoid being affected by the local optimal.The image compression based on GMO achieves an efficient coding compared to the stateof-the-art image coding algorithms.2.Considering the low quality of the predicted images,we propose an image compression based on Gaussian mixture model and Markov random field(GMM-MRF).The distribution of the priori probabilities of the Gaussian mixture model is described using Markov random field in the proposed method.The spatial constraints in the neighborhood of the image are introduced using Markov random field.The quality of the predicted image is improved in the proposed image compression based on GMM-MRF.Considering the high computational complexity of the image compression based on GMM-MRF,we propose Mixture Model Optimization based on K-Median algorithm.The proposed method calculates the MMO values using the initialized parameters provided by K-Median algorithm and selects the number of distributions corresponding to the minimum MMO as the optimal number of distributions.The number of distributions is selected before the estimation of the parameters,which reduces the computational complexity of the proposed image compression based on GMM-MRF.Because of the K-Median algorithm,MMO is not affected by the extreme data.MMO cannot avoid local optimum as GMO does,but MMO can effectively reduce the computational complexity of the propsoed coding algorithm.3.Considering the unclear similarity between Gaussian distributions in Gaussian mixture model,a decomposition algorithm of 3D covariance matrix based on the analysis of the regression plane is proposed in this paper.Firstly,by analyzing the relationship between the regression plane used to predict the gray value of pixels and the covariance parameters,the algorithm can obtain three vectors,which can represent the regression plane and recover the covariance parameters without errors.This method provides a new way to describe the geometric characteristics of a gaussian distribution.Secondly,according to the spatial relationships among the three vectors,the vectors are further transformed into five Gaussian features with clear physical meanings,which realizes the analysis of the complex three-dimensional covariance matrix.This method makes it possible to analyze the similarity between Gaussian distributions quantitatively.Utilizing the proposed Gaussian features,a feature-based dictionary can be established by matching the feature vectors of the Gaussian distributions.The matching of the features makes it possible to analyze the similarity between Gaussian distributions quantitatively.Based on the matching of the similar distributions,we can search and replace the similar areas in the elemental image array.The similarity between the elemental-images is effectively utilized in the proposed coding scheme.The proposed coding method,which achieves an efficient coding of elemental image arrays,outperforms the state-of-the-art coding methods.This paper has done some deep researches on image compression and coding of elemental image arrays based on Gaussian mixture model.The proposed image coding algorithms solve the problems of Gaussian mixture model when it is utilized in image coding.The proposed coding of elemental image arrays effectively utilizes the similarity between the elemental-images.This proposed coding scheme achieves an efficient coding of elemental image arrays and provides reference for further study on the GMM-based coding of elemental image arrays.