Unsupervised texture segmentation using resonance algorithm for natural scenes

https://doi.org/10.1016/S0167-8655(00)00035-0Get rights and content

Abstract

Many texture segmentation methods in the literature assume that the changes of intensity can be ascribed to the texture themselves. However, the real-world images may contain wide-ranged gradations in intensity which have nothing to do with local texture, such as those caused by the environment illuminations and cameras. To overcome the problem, an unsupervised texture segmentation method is proposed in this paper. Emphasizing the spatial relations between the adjacent texture pixels, the algorithm begins from a set of seed pixels and the texture region is generated by including those similar pixels. To suppress the noise influence, special attention is paid to the similarity criterion. Furthermore, to meet the requirement of unsupervised segmentation, the threshold in the similarity checking is automatically determined via iteratively applying the algorithm. The experimental results on Brodatz texture images and real-world images are presented.

Introduction

Texture segmentation plays an important role in both computer vision and image analysis. It consists of partitioning the input image into connected regions which are homogeneous with respect to a texture property. According to the methods used in the segmentation, it can be classified into supervised texture segmentation and unsupervised texture segmentation.

Unsupervised texture segmentation is a challenging task because of the lack of a prior knowledge about the texture in the image, such as how many different textures exist, what types of texture they are and where they are located. A lot of techniques for texture segmentation were proposed in the last two decades (Andrey and Tarroux, 1998, Tan, 1995, Jain and Farrokhnia, 1991, Chen and Pavlidis, 1979). The existing approaches are summarised by Reed and Buf (1993). Most approaches first map the differences in spatial structures, either stochastic or geometric, into the differences in feature space. This process is often called feature extraction. The performance of features is so closely related to the segmentation result that much research was devoted on this aspect (Buf et al., 1990, Shanmugam and Dinstein, 1973, Chen et al., 1995). Then, a segmentation algorithm is utilized on the feature space to extract the homogeneous regions. According to the segmentation methods, texture segmentation can be grouped into two classes: region-based and boundary-based. Region-based approach tries to identify regions which have a uniform texture while boundary-based approach attempts to detect the differences of texture in adjacent regions.

Although significant progress has been made during the last two decades, problems remain as the methods are sensitive to texture distortions. Many methods presented in the literature address the texture segmentation problem with the assumption that most of the change in intensity are ascribed to the textures themselves. Unfortunately, in reality, images may contain wide-ranged gradations in intensity which have nothing to do with local texture, such as those caused by the environment illuminations. Even though elaborate collections of lights can approximate uniform illumination of the scene, as a consequence of physical characteristics of lenses and cameras, vignetting still exists, resulting in the corners of the image being darker than the centre. Considering that the illumination varies gradually over the image and distortions are slight under most circumstance, we have developed an unsupervised texture segmentation method which overcomes the problem.

One way of coping with this problem is to use the features that are invariant to global illumination variations to some extent. Hild and Shirai (1993) extract texel properties, such as texel size, texel orientation and texel density as texture features for segmentation. Another way is to try to simulate the environment impacts and restore to a distortion-free image by imposing filters to the original picture. Chantler (1997) develops illumination tilt angle estimation in frequency domain while Knill (1990) develops in spatial domain. The success of the methods depends on the isotropic character of the surface texture. Russ (1992) suggests another method on nonuniform illumination by background subtraction, that is, composes an image simulating the background and subtracts/divides it from the original image.

For human vision, it is easy to distinguish different textures in spite of these fluctuations in the same texture. It seems that human beings do not explicitly deduce the influences of these variations but rather adapt to the gradient variations within the same texture. A similar strategy, resonance algorithm, is adopted in this paper. Unlike the above mentioned methods, it does not focus on the environment estimation but the feature variations in the neighbourhood region. In another word, it is supposed that the feature differences between adjacent neighbours of the same texture must be within a tolerable range. It differs from the widely-used clustering approaches in that it emphasises the geometrical locations as well as the feature differences. To meet the requirement of unsupervised segmentation, an automatic range selection method of the feature distances is suggested in this paper. Furthermore, a way to restrain the impacts of noise is proposed.

In Section 2, we discuss the mathematical description of the resonance algorithm. The noise suppression and threshold detection are also addressed in Section 2. The experiment results are presented in Section 3, followed by the summary.

Section snippets

Resonance algorithm

The algorithm can be described clearly via the resonance procedure in physical systems. Consider each pixel in the image as a mass connected to a platform through a spring and the whole image as a model composed of many such mass–spring pairs immersed in the water. Thus consequently, the mass–spring pairs are not isolated from each other but inter-connected through the water (Fig. 1). When we let an external force, for example, a sinusoidal force, act on the body, the behaviour of the

Experimental results

The whole segmentation scheme is shown in Fig. 5. The input image is tessellated into windows and the feature image is computed as the source of later processing. The features can be arbitrarily chosen from the characterized, vectorized data, such as features calculated from spatial gray level dependency, Gabor filter bank and wavelet methods, even illumination-dependent features. The following target is to label every pixel in the feature image and to map it back to the segmented image.

Summary

In this paper, an unsupervised texture segmentation method, resonance algorithm is proposed as it tolerates the texture distortions to some extent. Threshold is determined automatically via iteratively utilizing resonance algorithm to minimise the human intervention. Even the prior knowledge of the object number is unnecessary. To prevent the resonance spreading out from the noise pixels, noise suppression is used in the algorithm. Experimental results on Brodatz texture images and realworld

References (16)

There are more references available in the full text version of this article.

Cited by (8)

  • The application of resonance algorithm for image segmentation

    2007, Applied Mathematics and Computation
    Citation Excerpt :

    This is the influence of δ for the region segmentation, and comparatively, it can be used as criterion to estimate δ. Since the suitable threshold should be selected to ensure greater than the intra-region feature difference and less than the inter-region feature differences [1,7], from the initial seeds, the resonance begins from the current region to expend to other regions with the rise of δ. Fig. 4b gives the δ-area curves of Fig. 4a for the seed point selected in the top-left and bottom-right corner respectively, from which we see that the selection of seeds will affect the parameter δ, but not affect the segmentation result.

  • Area segmentation of images using watershed and anisotropic Gaussian kernels

    2019, International Journal of Pattern Recognition and Artificial Intelligence
  • Segmentation of Digital Images

    2017, Signals and Communication Technology
  • Unsupervised texture classification by combining multi-scale features and K-means classifier

    2009, Proceedings of the 2009 Chinese Conference on Pattern Recognition, CCPR 2009, and the 1st CJK Joint Workshop on Pattern Recognition, CJKPR
View all citing articles on Scopus
View full text