Lior Wolf, Xiaolei Huang, Ian Martin, Zhen Qian, Rui Huang, and Dimitris Metaxas
The detection of image edges has been one of the most explored domains in
computer vision. While most of the effort has been aimed at the detection of
intensity edges, the study of color edges and the study of texture edges are also
well developed fields.
The dominant approach in texture edge analysis is to construct a description
of the local neighborhood around each pixel, and then to compare this descriptor
to the descriptors of nearby points. This approach is often referred to as “patchbased”
since a fragment around each pixel is used in order to compute the
outputs of the filters. In this work, however, the term “patch-based” is quite
distinguishable from the above: it means that the gray values of the patch are
used as-is, and that the basic operation on patches is the comparison of two
patches using image correlation measures, such as normalized cross correlation
between the gray values, or their Euclidean distance.
What makes this approach novel for texture edge detection is that since
texture is a stochastic property, this kind of descriptor would traditionally be
The main idea of this work is simple to grasp: if a point lies on the left-hand side of a texture edge, the distribution of similarities of the patch centered at this point to the patches on its left is different from the distribution of similarities to the patches on its right. Detection of the texture edges can therefore be achieved by examining these differences in the similarity distributions. sampling from the distributions of similarities can be done very efficiently. In order to estimate whether the distributions are the same, we use a non-parametric test called the Wilcoxon Mann-Whitney Test. In our work, a novel technique for extracting texture edges is introduced. It is based on the combination of two ideas: the patch-based approach, and non-parametric tests of distributions. Our method can reliably detect texture edges using only local information. Therefore, it can be computed as a preprocessing step prior to segmentation, and can be very easily combined with parametric deformable models. These models furnish our system with smooth boundaries and globally salient structures.
We also present a deformable-model based solution for segmenting objects with complex texture patterns of all scales. The external image forces in traditional deformable models come primarily from edges or gradient information and it becomes problematic when the object surfaces have complex large-scale texture patterns that generate many local edges within a same region. We introduce a new textured object segmentation algorithm that has both the robustness of model-based approaches and the ability to deal with non-uniform textures of both small and large scales. The main contributions include an information-theoretical approach for computing the natural scale of a “texon” based on model-interior texture, a nonparametric texture statistics comparison technique and the determination of object belongingness through belief propagation. Another important property of the proposed algorithm is in that the texture statistics of an object of interest are learned online from evolving model interiors, requiring no other a priori information. We demonstrate the potential of this model-based framework for texture learning and segmentation using both natural and medical images with various textures of all scales and patterns.