Color–texture image segmentation by integrating directional operators into JSEG method
Introduction
Color image segmentation is one of the key processes in many applications based on image analysis, such as content-based image retrieval, image understanding, computer vision, and so on. It divides a color image into a set of disjoint regions which are homogeneous with respect to some properties, such as colors or textures. However, it is also a rather difficult problem to perform a perfect segmentation in color images.
In the past few decades, many methods (Cheng et al., 2001, Freixenet et al., 2002) have been proposed to segment the color image. Most of them were based on two basic properties of the pixels in relation to their local neighborhood: discontinuity and similarity. The approaches based on discontinuity attempted to partition an image by detecting isolated points, lines and edges, which were known as edge detection techniques. The simplest way to perform color segmentation may be to apply such classical gray-level gradient operators (Pratt, 1991) as Sobel, Prewitt, etc., in each color band of the image in a specified color space. On the other hand, region based approaches, including region growing, region splitting, region merging and their combination, have been also widely used. The well-known region growing method started with a seed region selected, and then grouped all similar neighbors according to a predefined homogeneity measure. Besides, to enhance the segmentation process, a large number of new algorithms (Freixenet et al., 2002, Fan et al., 2001, Xu and Shi, 2003) which integrated region and boundary information have been proposed over the last few years.
Segmentation can be also considered as an image classification problem based on color and spatial features. Therefore, we can categorize color segmentation methods as supervised or unsupervised learning/classification procedures. The JSEG method, newly proposed by Deng and Manjunath (2001), falls into the unsupervised class. The title of the algorithm came from the measure J which was defined as a criterion for a “good” segmentation. In the segmentation procedure, color similarity and spatial distributions of colors were taken into account separately to characterize differently homogeneous color–texture regions in the image. Although JSEG has been shown really robust on a variety of natural images, JSEG retains room for improvement. One potential limitation (Deng and Manjunath, 2001), as the authors of JSEG admitted, is that over-segmentation happens usually in case of spatially varying illumination.
Several variants have been proposed recently to fix JSEG’s inherent problems. Zheng et al. (2004) introduced fuzzy mechanism into JSEG to construct a soft class map for region growing, where each pixel was assigned into a set of classes with different membership degrees. It provided more detailed representation in labeling pixels’ classes and preserved the information of spatial distribution of colors better before region segmentation. Wang et al. (2004) also exploited the similar idea to avoid over-segmentation to some extent. Both of the above two methods enhanced JSEG from the aspect of the result of color quantization, i.e., the class map. Motivated from the homogeneity measure J, Jing et al. (2003) presented a new measure H for color image segmentation. Similarly, here we refer to this method as HSEG (H measure based SEGmentation). This heuristic criterion considered many factors, such as the symmetry, the periodicity and the scale of a pattern. Although HESG gave pretty good results on a number of natural images, we have found the H measure is more like edge detection operators (to be proven later) which is hard to describe textural homogeneity.
Our work focused on the pursuit of more effective homogeneity measures. In this paper, we firstly analyzed the advantages and disadvantages of the two measures J and H respectively and then improved the JSEG method by introducing a novel hybrid measure which integrated not only textural homogeneity but also color similarity. The rest of this paper is organized as follows. Section 2 reviews the two measures defined in JSEG and HSEG respectively and discusses the over-segmentation problem from a view of measure definition. Section 3 suggests a novel hybrid segmentation method which integrates color discontinuity obtained from directional operators into J measure. Comparative experiments are given in Section 4 and we conclude this paper in Section 5.
Section snippets
JSEG method
The basic idea of the JSEG method was to separate the segmentation process into two stages: color quantization and spatial segmentation. In the first stage, a color quantization algorithm based on peer group filtering (PGF) and vector quantization technique was adopted to reduce colors in the image. PGF was a nonlinear algorithm for image smoothing and impulse noise removal. It replaced each image pixel with the weighted average of its peer group members. The results of color quantization
A novel hybrid segmentation approach
The motivation of our work came from the above analysis on the advantages and disadvantages of the measures J and H. The integration of the two measures was expected to provide not only textural homogeneity characterized by J, but color discontinuity represented by H. Instead of the original single-band isotropic operators, we adopted the directional operators described as follows.
Experimental results
To evaluate the segmentation performance of the proposed method, we have carried out some experiments on the Berkley segmentation database (Martin et al., 2001) where the human segmented images provide our ground truth boundaries. The JSEG and HSEG algorithms have been also implemented for comparison. More than 150 color images are chosen randomly from this database and segmented by the three methods. No parameters are tuned on individual images to test the robustness of the algorithm. The same
Conclusions
A new homogeneity measure for color–texture image segmentation was proposed in this paper, which integrated both textural homogeneity and color discontinuity information to overcome the limitations of the JSEG and HSEG methods. The original measure J was defined over the class map without considering color discontinuity, which might result in inappropriate segmentation results. On the other hand, the measure H, although effective, was proven equivalent to the isotropic operators usually used
Acknowledgements
The authors are very grateful to the anonymous reviewers for their valuable suggestions.
References (12)
- et al.
Color image segmentation: Advances and prospects
Pattern Recognition
(2001) - et al.
Unsupervised segmentation of color-texture regions in images and video
IEEE Trans. Pattern Anal. Machine Intell.
(2001) - et al.
Automatic image segmentation by integrating color-edge extraction and seeded region growing
IEEE Trans. Image Process.
(2001) - et al.
Fast boundary detection: A generalized and a new algorithm
IEEE Trans. Comput.
(1997) - et al.
Yet another survey on image segmentation: Region and boundary information integration
Proc. ECCV02, LNCS
(2002) - et al.
Unsupervised image segmentation using local homogeneity analysis
Proc. IEEE Int. Symp. Circ. Sys.
(2003)
Cited by (31)
Registering georeferenced photos to a building information model to extract structures of interest
2019, Advanced Engineering InformaticsCitation Excerpt :As one of the most classical algorithms, JSEG was proposed by Deng and Manjunath [27] in 2001, which includes two steps, i.e. color quantization and spatial segmentation. Jing et al. [28] and Wang et al. [29], respectively, improved the JSEG algorithm by applying homogeneity analysis and combining directional operators. These methods have avoided human intervention for feature selection, but they often lead to over-segmentation, and fail to provide semantic information of the extracted ROI.
Image segmentation based on the integration of colourtexture descriptors - A review
2011, Pattern RecognitionCitation Excerpt :Implicit colour–texture feature integration [24–30,44]: the algorithms that belong to this category extract the texture features from single or multiple colour channels and the segmentation process is typically embedded into a coarse-to-fine strategy. Approaches that extract colour and texture in succession, i.e. the segmentation task is formulated as a sequence of serial processes [31,33–40,45,46,50,51]. Approaches that extract the colour and texture features on separate channels and then combine them in the segmentation process.
A Novel Approach for Quaternion Algebra Based JSEG Color Texture Segmentation
2023, International Journal on Recent and Innovation Trends in Computing and CommunicationUsing Adapted JSEG Algorithm with Fuzzy C Mean for Segmentation and Counting of White Blood Cell and Nucleus Images
2019, 2019 IEEE Asia-Pacific Conference on Computer Science and Data Engineering, CSDE 2019Automatic image annotation using possibilistic clustering algorithm
2019, International Journal of Fuzzy Logic and Intelligent Systems