Study On Texture Image Segmentation Using Non-negative Matrix Factorization And Active Contour Model

Image segmentation is a key technique for image processing,also a classic problem in temporary technique field.As an important visual image information,texture describes the space structures of grayscale in images,can be used for complex image segmentation.Comparing with the existing segmentation models,active contour models(ACMs)have been widely studied due to they can use smooth and closed curve to represent the segmentation results and provide sub-pixel accuracy.The basic idea of ACMs is to drive the active contours to move toward the object boundary by minimizing an energy functional.In existing ACMs for texture segmentation,there are following factors which cause the undesirable results: 1)the methods for describing image structures are relatively simple,which make segmentation models sensitive to clutter texture,noise and weak boundaries;2)for the estimation of the possibility that which region each pixel should belong to,this is usually done by using the distance between the representative features and each feature vectors.Due to the feature vectors computed from pixels around region boundaries usually consist of two or more region features.It is difficult to achieve an accurate boundary localization.Existing ACMs are usually established on the basis of the traditional level set framework.However,the recent studies have proved that this framework is easily coverage to a local optimum.To address these challenges,this paper presents a novel active contour model for accurate texture image segmentation.Some improvements focus on feature extraction,energy functional and convex optimization have been achieved,and more detailed works are described as follows:(1)To obtain more comprehensive texture description,a histogram based feature fusion strategy is presented to extract texture features.Firstly,the local variation degree of intensity(LVD)and Gabor filters are integrated together,to obtain multiple feature maps from input image.And then for each feature map,local histograms are computed over the fixed-size neighborhoods of each pixel.Finally,the feature histograms for each pixel is generated by concatenating all these histograms.This fusion strategy can improve the region discrimination and algorithmic robustness against noise and complex structures.(2)A novel method for computing the representative features of each region is proposed,named as Contour Shrinking Method.This method can prevent the feature histograms formed from multiple region features being used to compute the representative features of each region by moving the current contour inward and outward the object region.(3)To localize region boundaries more accurately,the Non-negative Matrix Factorization(NMF)method is embedded into our proposed energy functional.For each pixel,this NMF based energy can precisely approximate the coverage area each region has in its neighborhood,and to evaluate the probabilities of that pixel belongs to each region.Therefore,the similarity measurement between feature vectors and each representative feature can be naturally achieved.(4)Based on the convex optimization theory,the proposed ACM is converted to a convex minimization problem under the constraint that the minimizer remains fixed.After solving this optimization problem,the global minimizer of our proposed energy function can be achieved,and thus make our segmentation results away from local optimum.To validate the performance improvement brought by our model,the test experiments are conducted on varies kinds of images,including synthetic texture images,histology images and natural images.Moreover,some additional experiments are conducted to demonstrate the ability of our model on multi-object segmentation and the robustness against different initializations.From the segmentation results,we can conclude that our proposed method can obtain high accurate region boundaries in the presence of image noise and complex textures.