CN103258202B - A kind of texture characteristic extracting method of robust - Google Patents

Abstract

The invention discloses the texture characteristic extracting method of a kind of robust, belong to technical field of image processing.The step that realizes of the present invention is: input picture carries out pretreatment, generates feature set F, is then based on the threshold value of each feature by feature set F binaryzation, then carries out binary-coding generation specific pixel label；Input picture is rotated constant uniform LBP coding simultaneously, generates the LBP label of each pixel；Coexisted rectangular histogram by the specific pixel label of each pixel and LBP label configurations 2-D, be used for texture after the histogram vectors that will coexist and express.The application of the present invention, can reduce two-value in existing LBP mode and quantify loss, maintain simultaneously and extract the feature robustness to illumination, rotation, yardstick and visual angle change.

Description

A kind of texture characteristic extracting method of robust

Technical field

The invention belongs to technical field of image processing, particularly relate to the texture characteristic extracting method of a kind of robust.

Background technology

Textural characteristics plays important role in visual identity, has obtained widely studied and application in fields such as Texture classification, retrieval, synthesis and segmentations.It is said that in general, texture image not only presents various geometry and illumination variation, and change in being frequently accompanied by violent class and between class.When priori cannot obtain, Texture classification is a relatively difficult task.Therefore, the textural characteristics extracting robust is the key problem solving these tasks.

In the past few decades, the method that there has been proposed a lot of texture feature extraction.The research of early stage is focused mainly on different Statistics-Based Method, based on the feature with signal processing of model, such as co-occurrence matrix, Markov random field and the method etc. based on filtering.Later, the method based on primitive (textons) and local binary patterns (LBP) was suggested.The former needs a process learnt by the local feature of training image cluster is built primitive dictionary, then its primitive frequency of given texture statistics is set up rectangular histogram and expresses.The grey scale difference two-value quantization encoding of local, without training process, is directly expressed, is extracted the local microstructural information of texture with this by the latter.Method based on primitive is referred to document: M.VarmaandA.Zisserman, " Astatisticalapproachtomaterialclassifi-cationusingimagep atchexemplars; " IEEETrans.PatternAnal.Mach.Intell., vol.31, no.11, pp.2032 2047, Nov.2009；LBP is specifically referred to document: T.Ojala, M.Pietikainen, andT.Maenpaa, " Multiresolutiongray-scaleandrotationinvarianttextureclas sificationwithlocalbinarypatterns; " IEEETrans.PatternAnal.Mach.Intell., vol.24, no.7, pp.971 987, Jul.2002.

LBP utilizes a binary sequence to describe the feature of local grain, each pixel to image, the pixel value deducting this pixel with the pixel value of neighbour's sampling pixel points about obtains a corresponding sequence, take its value of symbol thus obtain one be encoded to 0/1 binary sequence, then this binary sequence is converted to the ten's digit Texture Identification as this pixel, i.e. the LBP code of this pixel.

LBP is simply efficient and famous with it, is widely used in the fields such as Texture classification, recognition of face and object detection.In recent years, people propose substantial amounts of innovatory algorithm based on LBP, wherein great majority are attributable to following several respects: the selection (such as ring-type or discoid geometry) of local pattern, sampling feature is (such as higher differentiation feature, Garbor feature, differential amplitude, block-based gray average), quantification manner (as three grades quantify and adaptive quantizing), coding rule is (such as division three value coding, statistical symbol number) and to high-dimensional extension (such as three-dimensional LBP, unifonn spherical region description, Garbor solid LBP and color LBP).

LBP is quantified as cost with two-value and exchanges the high efficiency extracting partial structurtes information and the poor robustness to noise for, quantifies loss to effectively reduce it, and maintains the robustness extracting feature, it is necessary to existing LBP mode is improved.

Summary of the invention

The goal of the invention of the present invention is in that: provide the texture characteristic extracting method of a kind of robust, to reduce two-value quantization loss in existing LBP mode, maintains simultaneously and extracts the feature robustness to illumination, rotation, yardstick and visual angle change.

The texture characteristic extracting method of a kind of robust of the present invention, comprises the following steps:

Step 1: generating specific pixel label L (x) of input picture I, wherein x represents the pixel of input picture I；

101: generate the n dimensional feature collection F={f of input picture I_i(x) | i=1,2 ..., n}；

102: described feature set F is carried out binary conversion treatment, obtain binary feature collection B={b_i(x) | i=1,2 ..., n；X ∈ I}:

If f_iX () is more than or equal to the threshold value thr of this feature_i, then b_iX () value is 1；It is otherwise 0；

103: described binary feature collection B is encoded, generates 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, rotates constant uniform LBP coding, generates LBP label Z (x) of each pixel input picture I；

Step 3, based on specific pixel label L (x) of each pixel, LBP label Z (x), structure 2 dimension symbiosis rectangular histogram, exports tie up symbiosis histogram vectors by described 2 after.

Original LBP method is texture feature extraction from the difference information of the little scope neighborhood of original image only, thus noise ratio is more sensitive.The present invention proposes a kind of texture blending method of robust, and the method realizes the extraction of textural characteristics based on the higher-order gradients domain information on bigger supporting zone, and the textural characteristics extracted than existing LBP mode robust more and has more expressiveness and judgement index.Relative to the texture characteristic extracting method based on primitive, the present invention omits its training taking a large amount of system resource and cluster process, directly the feature set of the input picture generated being carried out two-value quantization and coding, therefore the present invention is compared to the method based on primitive, and its realization is easier, efficient.

In order to obtain more stable quantization threshold, keep the high efficiency of texture feature extraction of the present invention, in a step 102, threshold value thr_iFor: each vegetarian refreshments characteristic of correspondence f on input picture I_iAverage on entire image I.

To sum up, the invention has the beneficial effects as follows, it is achieved easy, reduce two-value in existing LBP mode and quantify loss, maintain that to extract feature high to the robustness of illumination, rotation, yardstick and visual angle change and the textural characteristics judgement index that extracts simultaneously.

Accompanying drawing explanation

Examples of the present invention will be described by way of reference to the accompanying drawings, wherein:

Fig. 1 is the flow chart of the specific embodiment of the invention.

Detailed description of the invention

All features disclosed in this specification, or the step in disclosed all methods or process, except mutually exclusive feature and/or step, all can combine by any way.

This specification (include any accessory claim, summary and accompanying drawing) disclosed in any feature, unless specifically stated otherwise, all can by other equivalences or there is the alternative features of similar purpose replaced.That is, unless specifically stated otherwise, each feature is an example in a series of equivalence or similar characteristics.

In the texture feature extraction process of the present invention, the feature having more judgement power is obtained in order to reduce two-value to quantify loss, the present invention extracts useful information from bigger local neighbor supporting zone, maintain simultaneously and extract the feature robustness to illumination, rotation, yardstick and visual angle change, referring to Fig. 1, it implements step and is:

Step S100: generate the feature set F of input picture I, this step is the pre-treatment step of the present invention, implement the implementation that can adopt existing arbitrary maturation and all can realize the present invention, as in the present embodiment, adopted following step (1)～(4) to generate the feature set F of the present invention:

(1): input picture I is normalized, to remove illumination effect；Normalized can be existing arbitrary ripe mode, such as histogram equalization etc., it is preferred that be normalized by the average of image I and standard deviation；

(2): filter (Gauss second order local derviation) based on multiple dimensioned and multidirectional edge filter (Gauss single order local derviation) and strip and obtain the invariable rotary filter response of each yardstick respectively, namely respectively with the edge filter in m direction and the strip filtering normalized image I of convolution on each yardstick, and record the response of trend pass filtering that on this yardstick, amplitude response is maximum, so that each pixel x of image I obtains stable n-1 the dimension of feature set F (n represent) filter response；

The number of different yardsticks can set according to practical situation, preferably, considering 3 yardsticks, each yardstick adopts 8 trend pass filterings, and 3 yardsticks are taken as successively in the big I both horizontally and vertically gone up of image I: (1,3) (2,6), (3,12), can also according to the actual requirements and application, take other value.

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

(3) n-1 filter response of each pixel obtained being normalized, it is possible to be existing arbitrary normalization mode, the present invention preferably uses Web ' slaw (Weber's law) normalization, namely

R(x)←R(x)[log(1+M(x)/0.03)]/M(x)

Wherein, R (x) represents filter response, M (x)=| | R (x) | |₂Represent the amplitude of filter response, symbol | | | |₂Represent 2-norm.

(4): be made up of n normalized image I, n-1 filter response and tie up (n-D) feature set F, be expressed as:

F={f_i(x) | i=1,2 ..., n；X ∈ I}

Step S200: feature set F is carried out binarization operation, obtains binary feature collection B={b_i(x) | i=1,2 ..., n；X ∈ I}:

Wherein, thr_iRepresent feature f_iThreshold value, any feature f_iThe threshold value thr of (x)_iBoth can be based on empirical value to preset, it is preferred that set threshold value thr_iFor: each vegetarian refreshments characteristic of correspondence f on input picture I_iAverage on entire image I, namelyWherein X represents the number of the pixel x of input picture I.

Step S300: to binary feature collection B={b_i(x) | i=1,2 ..., n；X ∈ I} is encoded, and generates specific pixel tag L (x), namely

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

Step S400: input picture I is rotated constant uniform LBP coding, generates neighbor information code tag Z (x) of each pixel, namely

Wherein,Represent invariable rotary uniform pattern；Center pixel value g_cThe i.e. pixel value of current pixel point x；G_pBe center pixel surrounding sample radius it is the pixel value of pth the sampling pixel points of R；U (LBP_P,R) represent and uniformly estimate, the 0-1 transition times of the circumference bit string being namely made up of with the difference symbol of center pixel the pixel value of sampling neighbour；S (t) is sign function, and namely t is negative, and its functional value takes 0, otherwise takes 1.

Step S500: rectangular histogram is expressed.Based on specific pixel label L (x) of each pixel, LBP label Z (x), constructing 2-D symbiosis rectangular histogram, then its vectorization obtains a dimension is 2ⁿThe feature representation of × (P+2), this feature representation is final texture and expresses.2-D symbiosis histogram calculation mode is:

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

Wherein, (l p) represents that index is (l, 2-D rectangular histogram p), l ∈ [0,2 to Hⁿ-1], p ∈ [0, P+1],δ (y) is (l, number of times p) in order to pixel coder in accumulated image.Such as, for pixel x₀If, L (x₀)=1, Z (x₀)=2, then (l, in p), (l, p)=(1,2) corresponding value adds 1 to index to H.

In the present invention, the process generating specific pixel label L (x) and LBP label Z (x) both can executed in parallel, it is also possible to serial, selected according to practical application request.

With R=1, P=8,3 different scales (1,3), (2,6), (3,12) the upper edge filter adopting 8 directions and strip filtering, based on the texture characteristic extracting method of the present invention, it is used for there is illumination, rotates, the Outex of visual angle and dimensional variation, three texture database classification of CUReT and UIUC, contrast with the sorting algorithm based on LBP with based on study (such as primitive), and its classification performance is obviously improved；And the dimension of the feature representation finally given is 1280, compared to the method (typical dimension is 2440) based on primitive, it has relatively low characteristic dimension.

The invention is not limited in aforesaid detailed description of the invention.The present invention expands to any new feature disclosed in this manual or any new combination, and the step of the arbitrary new method disclosed or process or any new combination.

Claims (8)

1. the texture characteristic extracting method of a robust, it is characterised in that comprise the following steps:

Step 1: generating specific pixel label L (x) of input picture I, wherein x represents the pixel of input picture I；

101: generate the n dimensional feature collection F={f of input picture I_i(x) | i=1,2 ..., n}:

Input picture I is normalized；

Normalized image I is carried out multiple dimensioned Multi-aspect filtering: to each yardstick, adopt m trend pass filtering, obtain the maximum trend pass filtering of the amplitude response on each yardstick respectively as the invariable rotary filter response on this yardstick based on edge filter, strip filtering, and described filter response is normalized；

N dimensional feature collection F={f is constituted by normalized image I, filter response_i(x) | i=1,2 ..., n；X ∈ I}；

102: described feature set F is carried out binary conversion treatment, obtain binary feature collection B={b_i(x) | i=1,2 ..., n；X ∈ I}:

If f_iX () is more than or equal to the threshold value thr of this feature_i, then b_iX () value is 1；It is otherwise 0；

103: described binary feature collection B is encoded, generates specific pixel label L (x) of each pixel: $L\left(x\right)=\underset{i=1}{\overset{n}{\Σ}}{b}_i\left(x\right)2^{i-1};$

Step 2, rotates constant uniform LBP coding, generates LBP label Z (x) of each pixel input picture I；

Step 3, based on specific pixel label L (x) of each pixel, LBP label Z (x), structure 2 dimension symbiosis rectangular histogram, exports tie up symbiosis histogram vectors by described 2 after.

2. the method for claim 1, it is characterised in that in described step 102, described threshold value thr_iFor: each vegetarian refreshments characteristic of correspondence f on input picture I_iAverage on entire image I.

3. method as claimed in claim 1 or 2, it is characterised in that described normalized input picture I being normalized into removing light change.

4. method as claimed in claim 3, it is characterised in that the normalized of described removal light change is: be normalized by the pixel value average of each pixel of input picture I and standard deviation.

5. method as claimed in claim 1 or 2, it is characterized in that, described filter response is normalized into R (x) ← R (x) [log (1+M (x)/0.03)]/M (x), wherein R (x) represents filter response, M (x)=| | R (x) | |₂Represent the amplitude of filter response.

6. method as claimed in claim 1 or 2, it is characterised in that the m value of described m trend pass filtering is 8.

7. method as claimed in claim 1 or 2, it is characterised in that described multiple dimensioned be specially 3 yardsticks.

8. method as claimed in claim 7, it is characterised in that described 3 yardsticks are followed successively by the size both horizontally and vertically gone up of image I: (1,3), (2,6), (3,12).