Color Image Segmentation Based On Moment Classification And Statistical Modeling

With the continuous development of computer science and technology,digital image processing and analysis has gradually been a scientific system.It has become an effective tool for scholars to study visual perception in many fields such as psychology,physiology,and computer science.Image segmentation is the basic link of image processing and analysis.Its segmentation quality directly affects the subsequent operation effects,so it has been highly valued by people.However,there are some problems,such as poor performance when segmenting images which containing complex scenes,poor robustness,inaccurate description of image details and so on.In this paper,we propose three effective color image segmentation algorithms to deal with these problems,which can be summarized as follows:1?Based on the local description ability of faster and highly accurate quaternion exponential moments,a color image segmentation method against pixel classification is developed.First,the image is denoised,and window images are constructed which are centered on the pixel.Then,the image is decomposed using faster and highly accurate quaternion exponential moments.And the amplitude and relative phase of moments are calculated as pixel features.Finally,pixel features combined with the initial segmentation results of Tsallis entropy.And the final pixel classification is performed by the proximal classifier with consistency(PCC).In this paper,the faster and highly accurate quaternion exponential moments is presented,which can efficiently extract features of color images.Amplitude and relative phase are used to characterize features,which can improve the robustness of algorithm.Experimental results show that the proposed method provides an ability to describe and segment more accurately on color images.2?A color image segmentation method is proposed,which is based on multi-correlation hidden Markov tree(HMT)model of ECT domain.Its theoretical basis is empirical curvelet transform(ECT)and multivariate Cauchy statistical modeling.First,ECT is performed on R,G,and B channels respectively to obtain high frequency sub-bands.Then,ECT domain coefficients are statistically modeled by multivariate Cauchy distribution function,and the vector HMT model parameters are estimated.Finally,the data block likelihood values in different scales are calculated,and the target object is separated by text-based multi-scale fusion method.Three channel results are combined to obtain the final image segmentation result.In this paper,ECT is used to decompose the image,multivariate Cauchy distribution is used to fit the coefficient distribution,and vector HMT statistical modeling is introduced.The method fully considers the inter-scale and inter-direction correlation of coefficient sub-bands,and achieves accurate modeling.The experimental results show that the proposed method has higher segmentation accuracy.3?Combining hidden Markov random field(HMRF)and color significance(CS)theories,a fuzzy clustering segmentation method against color images is proposed.First,according to the saliency map of each channel,the comprehensive saliency map of image is calculated.Then,the closed arithmetic method of morphological reconstruction is used to remove the noise.It can effectively suppress noise and save image detail information.Finally,the fuzzy C-means clustering algorithm(FCM)based on membership degree filtering is used to process the HMRF models.The method fully considers the color information of color image by calculating the color information saliency map.And the membership function matrix is modified by using median filtering instead of the distance calculation between the neighborhood and the cluster center.This operation can increase the speed.Experimental results show that the proposed method has better segmentation performance.