texture features
Texture feature is a global feature and a visual feature that reflects homogeneity in an image. Image texture describes the spatial color distribution and light intensity distribution of an image or a small area in it, which reflects the slow-changing or periodic-changing surface tissue structure arrangement properties of the object surface. Image texture is represented by the gray distribution of pixels and their surrounding spatial neighborhoods, that is, local texture information. In addition, the repeatability of local texture information to varying degrees is global texture information.
Textures have three hallmarks:
-
A certain local sequence repeats itself;
-
non-random arrangement;
-
Roughly homogeneous unity within the textured area;
While the texture feature embodies the nature of the global feature, it also describes the surface properties of the scene corresponding to the image or image region. However, since texture is only a characteristic of the surface of an object, it cannot fully reflect the essential properties of the object, so it is impossible to obtain high-level image content only by using texture features. Different from color features, texture features are not pixel-based features, which require statistical calculations in regions containing multiple pixels. In pattern matching, this regional feature has great advantages, and it will not fail to match successfully due to local deviations.
Using texture features is an effective method when retrieving texture images with large differences in thickness and density. However, when there is little difference between textures, such as thickness and density, which are easy to distinguish, it is difficult for ordinary texture features to accurately reflect the differences between textures with different human visual perceptions. For example, reflections in water, the effect of smooth metal surfaces reflecting each other, etc. will all cause changes in texture. Since these are not properties of the object itself, when texture information is applied to retrieval, sometimes these false textures can cause "misleading" retrieval.
Characteristics of texture features
advantage:
-
Perform statistical calculations in areas containing multiple pixels;
-
Often invariant to rotation;
-
Strong resistance to noise;
shortcoming:
-
When the resolution of the image changes, the calculated texture may have a large deviation;
-
May be affected by light and reflection;
The texture reflected from the 2-D image is not necessarily the real texture of the surface of the 3-D object;
Texture Feature Classification
1. Basic instructions
The classification diagram of texture features is as follows:
the extraction of texture features is generally done by setting a window of a certain size, and then obtaining texture features from it. However, the choice of window has contradictory requirements:
The window setting is large: Texture is a regional concept, which must be reflected through spatial consistency. The larger the observation window is, the stronger the ability to detect the identity is; on the contrary, the weaker the ability is;
the window setting is small: because the boundary of different textures corresponds to the jump of the regional texture identity, in order to accurately locate the boundary, It is required to make the observation window smaller;
in this case, there will be difficulties: if the window is too small, there will be mis-segmentation inside the same texture; if the analysis window is too large, many mis-segments will appear in the texture boundary area.
When introducing the texture feature description method later, each method will be compared from the following four angles:
- Computational complexity
- Is it consistent with human visual experience
- Whether to use global information
- Whether it has multi-resolution characteristics
2. Texture feature description method
According to the texture feature description method, it can be divided into the following categories:
(1) Statistical method
The statistical method is based on the gray attribute of the pixel and its neighborhood to study the statistical characteristics of the texture area. The statistical characteristics include the first-order, second-order or higher-order statistical characteristics of the grayscale in the pixel and its neighborhood.
A typical representative of statistical methods is a texture analysis method called gray level co-occurrence matrix (GLCM). It is a method based on estimating the second-order combined conditional probability density of images. This method studies various statistical properties in the co-occurrence matrix through experiments, and finally obtains four key features in the gray-scale co-occurrence matrix: energy, inertia, entropy and correlation.
Although the texture features extracted by GLCM have good discriminative ability, this method is computationally expensive, especially for pixel-level texture classification. Moreover, the calculation of GLCM is time-consuming, but fortunately, researchers continue to propose improvements to it.
Other statistical methods include image autocorrelation function, semivariogram, etc.
advantage:
- The method is simple and easy to implement. In particular, the gray level co-occurrence matrix (GLCM) method is recognized as an effective method with strong adaptability and robustness;
shortcoming:
- It is out of touch with the human visual model, lacks the utilization of global information, and it is difficult to study the inheritance or dependence of pixels between texture scales;
- Lack of theoretical support;
- The computational complexity is high, which restricts the practical application.
(2) Geometric method
The geometric method is a texture feature analysis method based on the texture primitive theory, where the texture primitive is the basic texture element. Texture primitive theory holds that complex textures can be composed of several simple texture primitives arranged repeatedly in a certain regular form.
In the geometric method, the more influential algorithm is the Voronio checkerboard feature method.
However, the application and development of geometric methods are extremely limited, and there are few follow-up studies.
(3) Model method
There is an assumption in the model method: the texture is formed on the basis of a certain parameter-controlled distribution model.
Because the model method estimates and calculates the model parameters from the realization of the texture image, and uses the parameters as features, or uses a certain classification strategy for image segmentation, the estimation of the model parameters is the core issue of the model method.
Model-based texture feature extraction methods are mainly random field model methods and fractal model methods.
-
Random field model method: trying to describe the random process of texture with a probability model, they perform statistical operations on random data or random features, and then estimate the parameters of the texture model, and then cluster a series of model parameters to form and texture type number consistent model parameters. The estimated model parameters are used to estimate the maximum posterior probability of the gray image point by point, and determine the probability of the most likely belonging of the pixel in the case of the pixel and its neighborhood. Random field models actually describe the statistical dependence of pixels in an image on neighboring pixels.
-
Fractal model method: As an important feature and measure of fractal, fractal dimension combines the spatial information and gray information of the image simply and organically, so it has attracted people's attention in image processing. Research has shown that there is a very strong link between the human visual system's perception of roughness and bumpiness and the fractal dimension. Therefore, the texture characteristics of the image area can be described by the fractal dimension of the image area. The core problem of fractal dimension description texture is how to estimate fractal dimension accurately. The application of fractal dimension in image processing is based on two points:
(1) Different types of morphological substances in nature generally have different fractal dimensions;
(2) Due to the researchers' assumptions, there is a certain correspondence between fractals in nature and the grayscale representation of images.
Typical methods of random field model methods, such as Markov random field (MRF) model method, Gibbs random field model method, fractal model and autoregressive model.
advantage:
-
The method of the model family can take into account the local randomness of the texture and the overall regularity, and has great flexibility;
-
Using the random field model method to describe the texture features of remote sensing images and segment them on this basis, it conforms to or reflects the laws of geology to a large extent;
-
The main advantage of MRF is that it provides a general and natural model for expressing the interaction between spatially correlated random variables (it pays attention to the multi-resolution properties of texture, combined with the layering theory of images, develops the Hierarchical MRF methods, multi-resolution MRF methods, etc., can not only improve processing efficiency, but also study the inheritance or dependence of pixels between texture scales to obtain texture features).
shortcoming:
- Since the texture features are mainly identified by the model coefficients, it is difficult to solve the model coefficients;
- The calculation is very heavy. Since the texture image segmentation based on the MRF model is an iterative optimization process, its convergence speed from local to global is very slow (even if the conditional iterative mode (ICM) can speed up the search for solutions), it usually requires hundreds of iterations to converge;
- It is inconvenient to adjust the parameters, and the model should not be complicated.
(4) Signal processing method
The method of signal processing is based on time domain, frequency domain analysis, and multi-scale analysis. This method performs a certain transformation in a certain area of the texture image, and then extracts the eigenvalues that can remain relatively stable, and uses the eigenvalues as features to represent the consistency within the region and the dissimilarity between regions.
The texture features of the signal processing class mainly use some kind of linear transformation, filter or filter bank to transform the texture into the transform domain, and then apply some kind of energy criterion to extract the texture features. Therefore, methods based on signal processing are also called filtering methods. Most signal processing methods are proposed based on the assumption that the energy distribution in the frequency domain can identify textures.
The classic algorithms of signal processing methods include: gray level co-occurrence matrix, Tamura texture feature, autoregressive texture model, wavelet transform, etc.
advantage:
- Multi-resolution representation of textures enables analysis of textures on a finer scale;
- Wavelet conforms to human visual characteristics, and the extracted features are also beneficial to texture image segmentation;
- Ability to combine spatial/frequency domain analysis of texture features.
shortcoming:
- The multi-resolution decomposition of the orthogonal wavelet transform only further decomposes the low-frequency part, but does not consider the high-frequency part; and the texture information of real images often exists in the high-frequency part. Although wavelet packet analysis overcomes this shortcoming, it seems to be powerless to irregular textures; wavelets are mostly used in standard or regular texture images, but for natural images with more complex backgrounds, due to noise interference, or in a certain texture area Pixels are not similar everywhere, resulting in poor results of orthogonal wavelet transform;
- The amount of calculation is large.
(5) Structural analysis method
Structural analysis method believes that texture is described by the type and number of texture primitives, as well as the "repetitive" spatial organization structure and arrangement rules between primitives, and texture primitives are almost standardized. Relationship. Assuming that the primitives of the texture image can be separated, and the texture is segmented by primitive features and arrangement rules, it is obvious that the problem to be solved by the structural analysis method is to determine and extract the basic texture units, and to study the " Repeated" structure relationship.
Since the structural analysis method emphasizes the regularity of the texture, it is more suitable for analyzing artificial textures. However, a large number of natural textures in the real world are usually irregular. In addition, changes in decoupled strands are frequent, so the application of structural analysis is largely limited.
Typical algorithms of structural analysis: syntactic texture description algorithm, mathematical morphology method.
Summarize
To sum up, in terms of the effectiveness of extracting texture features, statistical methods, model methods, and signal processing methods are almost the same as geometric methods and structural analysis methods, and they have all been recognized.