Color texture image segmentation based on neutrosophic set and

  • View
    218

  • Download
    2

Embed Size (px)

Text of Color texture image segmentation based on neutrosophic set and

  • Computer Vision and Image Understanding 115 (2011) 11341144Contents lists available at ScienceDirect

    Computer Vision and Image Understanding

    journal homepage: www.elsevier .com/ locate /cviuColor texture image segmentation based on neutrosophic setand wavelet transformation

    Abdulkadir Sengur a, Yanhui Guo b,a Department of Computer and Electronics Science, Firat University, 23119 Elazig, Turkeyb Department of Radiology, University of Michigan, Ann Arbor, MI 48109, USA

    a r t i c l e i n f o a b s t r a c tArticle history:Received 21 April 2010Accepted 1 April 2011Available online 15 April 2011

    Keywords:Neutrosophic setColor texture image segmentationWavelet transform1077-3142/$ - see front matter 2011 Elsevier Inc. Adoi:10.1016/j.cviu.2011.04.001

    Corresponding author.E-mail address: yanhui.guo@aggiemail.usu.edu (Y.Efficient and effective image segmentation is an important task in computer vision and pattern recogni-tion. Since fully automatic image segmentation is usually very hard for natural images, interactiveschemes with a few simple user inputs are good solutions. In this paper, we propose a fully automaticnew approach for color texture image segmentation based on neutrosophic set (NS) and multiresolutionwavelet transformation. It aims to segment the natural scene images, in which the color and texture ofeach region does not have uniform statistical characteristics. The proposed approach combines colorinformation with the texture information on NS and wavelet domain for segmentation. At first, it trans-forms each color channel and the texture information of the input image into the NS domain indepen-dently. The entropy is defined and employed to evaluate the indeterminacy of the image in NSdomain. Two operations, a-mean and b-enhancement operations are proposed to reduce the indetermi-nacy. Finally, the proposed method is employed to perform image segmentation using a c-K-means clus-tering. The determination of the cluster number K is carried out with cluster validity analysis. Twodifferent segmentation evaluation criterions were used to determine the segmentations quality. Experi-ments are conducted on a variety of images, and the results are compared with those new existing seg-mentation algorithm. The experimental results demonstrate that the proposed approach can segment thecolor images automatically and effectively.

    2011 Elsevier Inc. All rights reserved.1. Introduction

    The task of image segmentation is to separate a given imageinto different regions with homogenous properties [1]. In a colortexture segmentation algorithm, a label is assigned to each pixelbased on its color properties and texture properties. Although sig-nificant works have focused on developing algorithms based oneither color [1,2] or texture features [3,4] separately, the area ofcombined color and texture segmentation remains open and ac-tive. Natural scenes generally contain the combination of colorand texture. Therefore, combining color and texture features wouldbe of significant benefit in distinguishing regions having the samecolor but different textures, and vice versa.

    In the past few decades, many methods have been proposed tosegment the color image [112]. Most of them were based on twobasic properties of the pixels in relation to their local neighbor-hood: discontinuity and similarity [1]. The approaches based ondiscontinuity partition an image by detecting isolated points, linesand edges, which were known as edge detection techniques [1517]. A particular class of segmentation techniques is based on mul-ll rights reserved.

    Guo).tiresolution analysis [7,10,1214]. Such methods are typicallybased on multiresolution transformations. In general, small detailsare detected in higher resolution images, while larger objects aresegmented in coarser images.

    Kim and Kim [5,6] proposed a multiresolution wavelet-basedwatershed image segmentation technique, using markers and a re-gion merging procedure to reduce over-segmentation. Jung andScharcanski [7] used a combined image denoising/enhancementtechnique based on a redundant wavelet transform for segmenta-tion. Ma and Manjunath [8,9] proposed the Edge Flow segmenta-tion technique, which consists of computing and updatingchanges in color and texture in a pre-defined scale. Deng and Manj-unath [11] proposed the JSEG method for multiscale segmentationof color and texture, based on color quantization and region grow-ing. Wy et al. [12] proposed a multiscale wavelet-based directionalimage force as external force for snake segmentation. Comaniciuand Meer [2] used a kernel in the joint spatial-range domain to fil-ter image pixels and a clustering method to retrieve segmented re-gions. Ozden and Polat [13] proposed a color image segmentationmethod based on low-level features including color, texture andspatial information. Chen et al. [14] proposed a color texture imagesegmentation algorithm based on wavelet transform and adaptiveclustering algorithm.

    http://dx.doi.org/10.1016/j.cviu.2011.04.001mailto:yanhui.guo@aggiemail.usu.eduhttp://dx.doi.org/10.1016/j.cviu.2011.04.001http://www.sciencedirect.com/science/journal/10773142http://www.elsevier.com/locate/cviu

  • A. Sengur, Y. Guo / Computer Vision and Image Understanding 115 (2011) 11341144 1135Arbelaez and Cohen [29] proposed an algorithm for constrainedsegmentation. The proposed method is a front propagation algo-rithm on the ultra-metric contour map that constructs Voronoi tes-sellations with respect to collections of subsets of the imagedomain. The algorithm is parameter-free, computationally efficientand robust. However, it needs to place the seed point inside eachobject of interest in the image by a human user.

    Ning et al. [30] presented a new region merging based interac-tive image segmentation method. It needs to roughly indicate thelocation and region of the object and background by using strokes,which are called markers. Moreover, a novel maximal-similaritybased region merging mechanism was proposed to guide the merg-ing process with the help of markers. The method is efficient butthe human user dependent.

    Zhang et al. [31] proposed a novel region-based ACM for imagesegmentation which was implemented with a new level set meth-od named Selective Binary and Gaussian Filtering Regularized Le-vel Set (SBGFRLS) method. The SBGFRLS method reduces theexpensive re-initialization of the traditional level set method tomake it more efficient.

    Although the above techniques present good results, severaldisadvantages of the reviewed methods still exist. For example,the mean-shift method requires a manual selection of spatial (hs)and color (hr) bandwidths, and optionally a minimum area param-eter (M) for region merging. Moreover, it is important to notice thatJSEG is intended to be an unsupervised segmentation method,meaning that it is free of user-defined parameter. On the otherhand, some of the methods such as [7,13,14] need no parameteradjustment; they yield over-segmentation and under segmenta-tion results.

    Neutrosophic set (NS), was proposed by Florentin Smarandacheas a new branch of philosophy dealing with the origin, nature andscope of neutralities, as well as their interactions with differentideational spectra [18]. In neutrosophy theory, every event hasnot only a certain degree of truth, but also a falsity degree andan indeterminacy degree that have to be considered independentlyfrom each other [18]. Thus, a theory, event, concept, or entity, {A} isconsidered with its opposite {Anti-A} and the neutrality {Neut-A}.{Neut-A} is neither {A} nor {Anti-A}. The {Neut-A} and {Anti-A} arereferred to as {Non-A}. According to this theory, every idea {A} isneutralized and balanced by {Anti-A} and {Non-A} [18]. NS pro-vides a powerful tool to deal with the indeterminacy. Guo andCheng gave an example about reviewing paper to explain howthe NS works and how better it is [24].

    Nowadays, neutrosophic algebraic structures have been investi-gated [19], while the neutrosophic framework has found severalpractical applications in a variety of different fields, such as rela-tional data base systems, semantic web services [18], financialdataset detection [19] and new economies growth and declineanalysis [20]. In this paper, we propose a new approach for colortexture image segmentation that is based on neutrosophic set(NS) and multiresolution wavelet decompositions. We combinecolor information with the low-level features of grayscale compo-nent of the texture. The proposed approach transforms both eachcolor channel of the input image and the wavelet decompositionof the grayscale image into the NS domain independently whichis described using three membership sets: T, I and F. The entropyin NS is defined and employed to evaluate the indeterminacy.Two operations, a-mean and b-enhancement operations are pro-posed to reduce the set indeterminacy. Finally, the proposed meth-od is employed to perform image segmentation using a c-meansclustering. Experiments are conducted on a variety of images,and our results are compared with those new existing segmenta-tion algorithm.

    In the proposed method, it overcomes the drawback of the ori-ginal NS based segmentation method, and NS is used to describethe indeterminacy in the color texture image. The experimental re-sults demonstrate that the proposed approach can segment thecolor images automatically and effectively.

    In summary, the novelties of this paper are the following:

    A fully automatic approach for color texture image segmenta-tion based on neutrosophic set and multiresolution wavelettransformation that enables segment the color texture imagewithout human inactivation. Adaptively selection on parameters in neutrosophic set that

    enables to reduce the indeterminacy according to the character-istic of the image. Cluster validity analysis to determination of the cluster number.