CN103258202A - Method for extracting textural features of robust - Google Patents

Abstract

The invention discloses a method for extracting textural features of a robust, and belongs to the technical field of image processing. The method comprises the implementation steps: pre-processing an input image, generating a feature set F, carrying out binaryzation on the feature set F based on a threshold value of each feature, carrying out binary encoding and generating a particular pixel label, carrying out rotation invariant even local binary pattern (LBP) encoding on the input image, generating an LBP label of each pixel point, constructing a 2-D coexistence histogram by the particular pixel label and the LBP label of each pixel point, vectoring the coexistence histogram and then applying the coexistence histogram into textural expression. The method is applied so as to reduce binary quantitative loss in an existing LBP mode, at the same time maintain robustness of changes of illumination, rotation, scale and visual angles by the extraction features.

Description

Robust texture feature extraction method

Technical Field

The invention belongs to the technical field of image processing, and particularly relates to a robust texture feature extraction method.

Background

Texture features play an important role in visual recognition, and have been widely researched and applied in the fields of texture classification, retrieval, synthesis, segmentation, and the like. In general, texture images not only exhibit a wide variety of geometric and lighting changes, but are often accompanied by drastic intra-class and inter-class variations. Texture classification is a difficult task when a priori knowledge is not available. Therefore, extracting robust texture features is a core problem to solve these tasks.

Over the past several decades, many methods have been proposed to extract textural features. Early research focused on different statistical-based methods, model-based and signal-processing features such as co-occurrence matrices, markov random fields and filter-based methods, etc. Later, methods based on primitives (textons) and Local Binary Patterns (LBP) were proposed. The former requires a learning process to build a primitive dictionary by clustering local features of the training images, and then to build a histogram expression by counting the primitive frequencies for a given texture. The local gray difference binary quantization coding is directly expressed by local gray difference binary quantization coding without a training process, so that local microstructure information of the texture is extracted. Examples of primitive-based methods are found in M.Varma and A.Zisserman, "A statistical adaptation to material class-using image patch templates," IEEE trans.Pattern animal. inner. vol.31, No.11, pp.2032-2047, Nov.2009; LBP is specifically referenced: t.ojala, m.pietikaine, and t.maenpaa, "Multiresolution grid-scale and rotation in variable temporal mapping with local binary patterns," IEEE trans.pattern anal.mach.intell., vol.24, No.7, pp.971-987, jul.2002.

The LBP uses a binary sequence to describe the characteristics of local texture, for each pixel point of the image, the pixel value of the pixel point is subtracted from the pixel value of the neighboring sampling pixel point around the pixel point to obtain a corresponding sequence, the symbol value of the corresponding sequence is taken to obtain a binary sequence coded as 0/1, and then the binary sequence is converted into decimal numbers to be used as the texture identification of the pixel point, namely the LBP code of the pixel point.

LBP is well known for its simplicity and efficiency, and has been widely used in the fields of texture classification, face recognition, and object detection. In recent years, a number of improved algorithms have been proposed based on LBP, most of which are attributable to the following: selection of local patterns (e.g., ring or disk geometry), sampling features (e.g., high order differential features, Garbor features, differential amplitude, block-based gray-scale mean), quantization approaches (e.g., three-level quantization and adaptive quantization), coding rules (e.g., split ternary coding, statistical sign number), and extensions to high dimensions (e.g., stereo LBP, uniform spherical region description, Garbor stereo LBP, and color LBP).

LBP trades binary quantization for high efficiency of extracting local structure information and poor robustness to noise, and in order to effectively reduce quantization loss and maintain robustness of extracted features, it is necessary to improve the current LBP method.

Disclosure of Invention

The invention aims to: a robust texture feature extraction method is provided to reduce binary quantization loss in the existing LBP mode, and meanwhile, robustness of extracted features to illumination, rotation, scale and view angle changes is maintained.

The invention relates to a robust texture feature extraction method, which comprises the following steps:

step 1: generating a specific pixel label L (x) of the input image I, wherein x represents a pixel point of the input image I;

101: generating an n-dimensional feature set F ═ F for an input image I_i(x)|i＝1,2,...,n}；

102: and carrying out binarization processing on the feature set F to obtain a binary feature set B ═ B_i(x)|i＝1,2,...,n;x∈I}：

If f_i(x) Greater than or equal to the threshold value thr of the feature_iThen b is_i(x) The value is 1; otherwise, the value is 0;

103: encoding the binary feature set B to generate a specific pixel label L (x) of each pixel:

$L\left(x\right)=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{b}_i\left(x\right)2^{i-1};$

step 2, carrying out rotation invariant uniform LBP coding on the input image I to generate an LBP label Z (x) of each pixel point;

and 3, constructing a 2-dimensional symbiotic histogram based on the specific pixel label L (x) and the LBP label Z (x) of each pixel point, and vectorizing and outputting the 2-dimensional symbiotic histogram.

The original LBP method extracts texture features from only the difference information of a small-range neighborhood of the original image, and is sensitive to noise. The invention provides a robust texture extraction method, which realizes extraction of texture features based on high-order gradient domain information on a larger support region, and the extracted texture features are more robust and have expression force and discrimination force compared with the conventional LBP (local binary pattern) mode. Compared with the texture feature extraction method based on the primitive, the method omits the training and clustering processes which occupy a large amount of system resources, and directly performs binary quantization and coding on the feature set of the generated input image, so that the method is simpler, more convenient and more efficient to realize compared with the method based on the primitive.

In order to obtain a relatively stable quantization threshold and maintain the high efficiency of the present invention in extracting texture features, in step 102, a threshold value thr_iComprises the following steps: characteristic f corresponding to each pixel point on input image I_iMean over the whole image I.

In conclusion, the method has the advantages of being simple and convenient to implement, reducing binary quantization loss in the conventional LBP mode, maintaining robustness of extracted features on illumination, rotation, scale and visual angle change and enabling the extracted texture features to have high discriminability.

Drawings

The invention will now be described, by way of example, with reference to the accompanying drawings, in which:

FIG. 1 is a flow chart of an embodiment of the present invention.

Detailed Description

All of the features disclosed in this specification, or all of the steps in any method or process so disclosed, may be combined in any combination, except combinations of features and/or steps that are mutually exclusive.

Any feature disclosed in this specification (including any accompanying claims, abstract and drawings), may be replaced by alternative features serving equivalent or similar purposes, unless expressly stated otherwise. That is, unless expressly stated otherwise, each feature is only an example of a generic series of equivalent or similar features.

In the texture feature extraction process of the present invention, in order to reduce the binary quantization loss and obtain the feature with more judgment power, the present invention extracts useful information from a larger local neighbor support region, and simultaneously maintains the robustness of the extracted feature to illumination, rotation, scale and view angle changes, see fig. 1, and the specific implementation steps thereof are:

step S100: the feature set F of the input image I is generated, which is a preprocessing step of the present invention, and the present invention can be implemented in any existing mature implementation manner, for example, in the present embodiment, the following steps (1) to (4) are adopted to generate the feature set F of the present invention:

(1): normalizing the input image I to remove the illumination influence; the normalization process can be any existing mature mode, such as histogram equalization and the like, and is preferably performed by using the mean value and standard deviation of the image I;

(2): respectively obtaining rotation invariant filter responses of all scales based on multi-scale and multi-direction edge filtering (first order Gaussian partial derivative) and strip filtering (second order Gaussian partial derivative), namely, convolutively normalizing the image I by using edge filtering and strip filtering in m directions on each scale, and recording the direction filter response with the maximum amplitude response on the scale, so that each pixel point x of the image I obtains stable n-1 (n represents the dimension of a feature set F) filter responses;

the number of different scales can be set according to actual conditions, preferably, 3 scales are considered, each scale adopts 8-direction filtering, and the sizes of the 3 scales in the horizontal and vertical directions of the image I can be sequentially selected as follows: (1, 3) (2, 6), (3, 12), and other values can be taken according to actual requirements and applications.

When based on 3 scales, each pixel point x will obtain 6 stable filter responses.

(3) The normalization processing is carried out on the obtained n-1 filter responses of each pixel point, which can be any one of the existing normalization modes, and the invention preferably adopts Web's law (Weber's law) normalization, namely

R(x)←R(x)[log(1+M(x)/0.03)]/M(x)

Wherein r (x) represents a filter response, m (x) | | | r (x) | luminance₂Representing the magnitude of the filter response, the symbol | · | | non-woven calculation₂Representing a 2-norm.

(4): an n-dimensional (n-D) feature set F is constructed from the normalized image I, n-1 filter responses, and is expressed as:

F＝{f_i(x)|i＝1,2,...,n；x∈I}

step S200: carrying out binarization operation on the feature set F to obtain a binary feature set B ═ B_i(x)|i＝1,2,...,n;x∈I}：

Wherein, thr_iRepresenting a feature f_iThreshold value of, any feature f_i(x) Threshold value thr of_iThe threshold value thr may be preset, preferably, based on an empirical value_iComprises the following steps: characteristic f corresponding to each pixel point on input image I_iMean over the entire image I, i.e.

Wherein X represents the number of pixel points X of the input image I.

Step S300: for the binary feature set B ═ B_i(x) I1, 2, n, x belongs to I, and generates a specific pixel label L (x), namely

$L\left(x\right)=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{b}_i\left(x\right)2^{i-1}$

Step S400: carrying out rotation invariant uniform LBP coding on an input image I to generate a neighbor information coding label Z (x) of each pixel, namely

Wherein,

indicating a rotationally invariant uniform pattern; central pixel value g_cNamely the pixel value of the current pixel point x; g_pThe pixel value of the p-th sampling pixel point with the sampling radius of R around the central pixel is obtained; u (LBP)_P,R) Representing the number of 0-1 transitions of a uniform measure, i.e. a circumferential bit string consisting of the sign of the difference between the pixel values of the sampling neighbours and the central pixel; s (t) is a sign function, i.e., t is a negative number, and the function value is 0, otherwise it is 1.

Step S500: histogram expression. Constructing a 2-D symbiotic histogram based on specific pixel labels L (x) and LBP labels Z (x) of all pixel points, and then vectorizing the symbiotic histogram to obtain a 2-D symbiotic histogram with dimension of 2ⁿX (P +2), which is the final texture expression. The 2-D symbiotic histogram calculation mode is as follows:

$H(l,p)=\underset{x}{\mathrm{\Σ}}\mathrm{\δ}((L\left(x\right),Z\left(x\right))==(l,p)$

where H (l, p) represents a 2-D histogram with an index of (l, p), l ∈ [0,2 ]ⁿ-1]，p∈[0,P+1]，

δ (y) is used to accumulate the number of times a pixel in the image is encoded as (l, p). For example, for pixel x₀If L (x)₀)＝1，Z(x₀) When the value is 2, 1 is added to the value corresponding to the index (l, p) = (1,2) in H (l, p).

In the present invention, the processing for generating the specific pixel label l (x) and the LBP label z (x) may be executed in parallel or in series, and is selected according to the actual application requirements.

The texture feature extraction method is used for classifying three texture databases of Outex, CURET and UIUC with illumination, rotation, visual angle and scale change by adopting edge filtering and strip filtering in 8 directions on 3 different scales (1, 3) (2, 6) and (3, 12) with R =1 and P =8, and compared with classification algorithms based on LBP and learning (such as primitives), the classification performance is obviously improved; and the resulting feature representation has a dimension of 1280, which has a lower feature dimension compared to primitive-based methods (typical dimension is 2440).

The invention is not limited to the foregoing embodiments. The invention extends to any novel feature or any novel combination of features disclosed in this specification and any novel method or process steps or any novel combination of features disclosed.

Claims (9)

1. A robust texture feature extraction method is characterized by comprising the following steps:

step 1: generating a specific pixel label L (x) of the input image I, wherein x represents a pixel point of the input image I;

101: generating an n-dimensional feature set F ═ F for an input image I_i(x)|i＝1,2,...,n}；

102: and carrying out binarization processing on the feature set F to obtain a binary feature set B ═ B_i(x)|i＝1,2,...,n;x∈I}：

If f_i(x) Is greater than or equal toThreshold value thr of characteristic_iThen b is_i(x) The value is 1; otherwise, the value is 0;

103: encoding the binary feature set B to generate a specific pixel label L (x) of each pixel:

$L\left(x\right)=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{b}_i\left(x\right)2^{i-1};$

step 2, carrying out rotation invariant uniform LBP coding on the input image I to generate an LBP label Z (x) of each pixel point;

and 3, constructing a 2-dimensional symbiotic histogram based on the specific pixel label L (x) and the LBP label Z (x) of each pixel point, and vectorizing and outputting the 2-dimensional symbiotic histogram.

2. The method of claim 1, wherein the threshold value thr in step 102_iComprises the following steps: characteristic f corresponding to each pixel point on input image I_iMean over the whole image I.

3. The method according to claim 1 or 2, wherein in step 101, generating the n-dimensional feature set F is specifically:

carrying out normalization processing on the input image I;

performing multi-scale multi-directional filtering on the normalized image I: for each scale, m directional filtering is adopted, the directional filtering with the maximum amplitude response on each scale is respectively obtained based on edge filtering and strip filtering to serve as the rotation invariant filtering response on the scale, and normalization processing is carried out on the filtering response;

an n-dimensional feature set F is formed by the normalized image I and the filter response: f ═ F_i(x)|i＝1,2,...,n；x∈I}。

4. The method according to claim 3, wherein the normalizing process on the input image I is a normalization process that removes an illumination transform.

5. The method of claim 4, wherein the normalization process to remove the illumination transform is: and carrying out normalization processing by using the pixel value mean value and the standard deviation of each pixel point of the input image I.

6. The method of claim 3, wherein normalizing the filter response is: r (x) ← R (x) [ log (1+ M (x)/0.03)](x), where r (x) represents a filter response, m (x) r (x) y₂Representing the magnitude of the filter response.

7. The method of claim 3, wherein m of the m directional filters is 8.

8. The method of claim 3, wherein the multi-scale is specifically 3-scales.

9. The method of claim 8, wherein the 3 dimensions have magnitudes in the horizontal and vertical directions of the image I in the order: (1,3)(2,6),(3,12).