CN103955691A - Multi-resolution LBP textural feature extracting method - Google Patents
Multi-resolution LBP textural feature extracting method Download PDFInfo
- Publication number
- CN103955691A CN103955691A CN201410191280.9A CN201410191280A CN103955691A CN 103955691 A CN103955691 A CN 103955691A CN 201410191280 A CN201410191280 A CN 201410191280A CN 103955691 A CN103955691 A CN 103955691A
- Authority
- CN
- China
- Prior art keywords
- lbp
- image
- frequency
- histogram
- multiresolution
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Landscapes
- Image Analysis (AREA)
Abstract
The invention discloses a multi-resolution LBP (short for MR-LBR) textural feature extracting method, and belongs to the technical field of image information processing. The method includes the steps that firstly, an input image is preprocessed; secondly, image signals are decomposed and expressed through first-level discrete wavelet transformation, and a high-frequency mean chart is acquired; thirdly, rotation invariant unified LBP calculation is conducted on an original image, a low-frequency approximate subgraph and the high-frequency mean chart, and then an LBP histogram of the original image, an LBP histogram of the low-frequency similar subgraph and an LBP histogram of the high-frequency mean chart are acquired; eventually, the three histograms are spliced into a multi-resolution LBP histogram in a non-overlapping mode for describing texture information of the image. By the application of the method, more textural feature information of the image can be extracted, the defects of existing LBP in the texture processing aspect are overcome, and robustness of extracted features on rotation, illumination and the like is maintained. The multi-resolution LBP textural feature extracting method is applied to classification of fresh green tea leaves, the classification effect is remarkable, and accuracy reaches up to over 92 %.
Description
Technical field
The invention belongs to technical field of image information processing, relate in particular to the texture characteristic extracting method (MR-LBP) of a kind of multiresolution LBP.
Background technology
Texture is a kind of important visual cues, and it is that people describe and distinguish one of key character of different objects.Texture analysis is being played the part of very important role in fields such as Images Classification, image retrieval and industrial detection.Generally speaking, image texture is easily subject to the impact that illumination and rotation etc. change, and in the time cannot extracting more exactly the texture statement of image, it is more difficult that Texture classification work just seems.Therefore, be necessary to propose a kind of new texture characteristic extracting method to solve these key problems.
Now there is the method for many texture feature extraction, for example: size extraneous features conversion (SIFT), gray level co-occurrence matrixes (GLCM gray level co-occurrence matrix), wavelet transformation and Gabor filtering etc.In recent years, the local binary pattern (LBP local binary patterns) that Ojala et al. (2002) proposes has been obtained larger success, because its principle is relatively simple, computation complexity is low, there is again the remarkable advantage such as rotational invariance and gray scale unchangeability, thereby the method is widely used in the field such as detection and tracking and biomedical image analysis of images match, moving target simultaneously.
LBP operator utilizes a binary sequence Description Image Local textural feature.Its basic thought is: the rectangular block of choosing fixed window size (being commonly defined as 3 × 3) extracts primitive information, as threshold value, its neighborhood sub-block is carried out respectively to binary conversion treatment (if the pixel value in neighborhood is less than the gray-scale value of center pixel by center sub-block, this location of pixels is designated as 0, otherwise be designated as 1), then be multiplied by corresponding weight according to different positions, finally summation obtains the LBP value of this texture cell.
Although LBP is with simply efficiently famous, still there is part limitation in it, for example: have to noise and illumination variation sensitivity the problem such as histogram is bigger than normal.In order to address the above problem, researchist has made many improvement to LBP, such as: the local binary patterns (Extended LBP) of the expansion that soft histogram (Soft Histograms) method and the Zhou etc. that Bayes's local binary patterns (Bayesian LBP), the Ahone etc. that He etc. propose proposes proposes etc., improve to a certain extent LBP to the robustness of disturbing although these are improved one's methods, fundamentally do not solved the instability that threshold process is brought.In addition, directly on gray level image, calculate LBP feature and also can introduce much noise information, and the feature that only adopts a kind of LBP operator to extract is very limited.Meanwhile, existing LBP mode is not good enough to the relatively chaotic large-size images treatment effect of texture.
In the process of image texture classification, the local feature of image and all play a part of equal importance with global characteristics.So introduce wavelet transformation, image is carried out to multiresolution analysis, can take into account the information representation of entire and part simultaneously, thereby be unlikely to cause the dominant character of image to be lost or inadequate obtaining.
Summary of the invention
Goal of the invention of the present invention is: the texture characteristic extracting method (MR-LBP) that a kind of multiresolution LBP is provided, loss to image section information while quantification to solve traditional LBP operator two-value, improve its deficiency in the time processing irregular, the larger-size complicated image of texture simultaneously, and maintain and extract the robustness of feature to rotation and illumination etc.
The texture characteristic extracting method of a kind of multiresolution LBP that the present invention proposes, comprise the following steps: that (1) generates four subbands of input picture I on different frequency domains, comprising a low frequency sub-band (low-frequency approximation subgraph) and three high-frequency sub-band (high frequency detail view), be to obtain through one-level wavelet transform (DWT) to four described subbands, wavelet transformation can provide multiresolution analysis for image; (2) obtain after wavelet decomposition three high-frequency sub-band are averaged, generate a high-frequency average figure
; (3) original image, low-frequency approximation subgraph and high-frequency average figure are rotated respectively to constant unified LBP and calculate, and obtain by statistics LBP histogram H (i) separately.Invariable rotary is unified LBP operator
, be the result of invariable rotary LBP and " More General Form " combination, the dimension of described LBP histogram H (i) is 10; (4) construct 30 dimension multiresolution LBP histograms based on three LBP histograms, will after described 30 dimension multi-resolution histogram vectorizations, export.
Original LBP operator directly acts on gray level image and extracts eigenwert, thereby can introduce a large amount of noise informations.In addition LBP processing that texture is relatively chaotic, when larger-size complicated image effect not good enough.The present invention proposes the texture characteristic extracting method of a kind of multiresolution LBP, the method can be from three different resolution, extract respectively the eigenwert of image, then be spliced into a comprehensive proper vector, can make like this texture description of image more perfect, for Texture classification provides favourable data basis.With respect to LBP and wavelet transformation itself, the method that the present invention proposes has not only taken into full account the global information (being low frequency sub-band information) of texture image, and special concern the detailed information of texture image (being high-frequency sub-band information).Therefore the present invention can more fully explain image texture information, and realizes fairly simplely, has more generality.
To sum up, the invention has the beneficial effects as follows, realize fairly simplely, can take into account the information representation of entire and part simultaneously, thereby be unlikely to cause the dominant character of image to be lost or inadequate obtaining.And can effectively improve the deficiency of existing LBP in the time processing irregular, the larger-size complicated image of texture.Also maintain and extract the robustness of feature to rotation and illumination simultaneously.
Brief description of the drawings
Invention will be explained by example and with reference to the mode of accompanying drawing.
Fig. 1 is the process flow diagram of the specific embodiment of the invention.
Fig. 2 is many resolution LBP histograms that method disclosed by the invention is extracted.
Embodiment
Disclosed all features in this instructions, or step in disclosed all methods or process, except special circumstances, all can be combined with any mode.
[in texture feature extraction process of the present invention, in order to obtain the more real textural characteristics of image ratio, first the present invention has carried out one-level wavelet transform on to the pretreated basis of image, and image is carried out to multiresolution analysis.Also maintained the stability of extracting feature rotation and illumination etc., referring to Fig. 1, its specific implementation step is: step S100: generate four subbands of input picture I on different frequency domains, this step is carried out on the pretreated basis of image simultaneously.Adopt following step to become four subbands of the present invention next life: (1) carries out pre-service to image, to remove the impact of the factors such as illumination.(2) select Haar wavelet basis to carry out the decomposition of one-level discrete wavelet to image, suppose to have the original texture image of N × M size, pixel gray-scale value is
, after wavelet transform, can be described as:
(1)
(2)
(3)
(4)
After S level is decomposed, can obtain (3S+1) individual subgraph:
, in the time of S=1, can obtain (3*1+1)=4 subgraph, now K=1.Wherein
represent original tealeaves image
the low frequency component of horizontal and vertical direction;
represent that scale factor is
horizontal direction low frequency component and vertical direction high fdrequency component;
represent that scale factor is
horizontal direction high fdrequency component and vertical direction low frequency component;
represent that scale factor is
the high fdrequency component of horizontal and vertical direction; Step S200: obtain after wavelet decomposition three high-frequency sub-band are averaged, generate a high-frequency average figure
, specific implementation is:
, after averaging by the signal on three high-frequency sub-band correspondence positions, obtain a high-frequency average figure; Step S300: original image, low-frequency approximation subgraph and high-frequency average figure are rotated respectively to constant unified LBP and calculate, and obtain by statistics LBP histogram H (i) separately, that is:
, wherein
represent invariable rotary More General Form; P refers to the number of pixels of subwindow center pixel neighborhood, and R refers to the radius taking center pixel as center of circle neighborhood; f
crefer to the gray-scale value of center pixel, f
p(p=0,1 ..., 7) and refer to the gray-scale value of 8 pixels in its neighborhood.U (LBP
p,R) represent Unified Measure, the corresponding circulation binary number from 0 to 1 or from 1 to 0 of certain local binary pattern has at most twice saltus step.U (t) is-symbol function, when t is non-when negative, value is 1, otherwise is 0.Unify after LBP calculates to obtain the value between a group 0 ~ 9 through invariable rotary, add up number corresponding to each value and can obtain LBP histogram; Step S400: the LBP histogram that is 10 by three dimensions that obtain in step S300 is not spliced into the multiresolution LBP histogram of one 30 dimension overlappingly, will export after described 30 dimension multi-resolution histogram vectorizations.H (I)=H (
), H (
), H (
), wherein H (I) represents the multiresolution LBP histogram of piece image, H (
), H (
) and H (
) represent respectively the LBP histogram of original image, low-frequency approximation subgraph and high-frequency average figure.
In the present invention, three histogrammic processing procedures of LBP can be according to the needs of concrete practical application, select flexibly serial/parallel to carry out.
Invariable rotary taking P=8, R=1 as parameter is unified LBP operator and is processed image.Based on texture characteristic extracting method of the present invention, image in VisTex and two data texturing storehouses of the fresh leaf of green tea is carried out respectively to classification experiments: in two experiments, first we extract the textural characteristics of each image, then the input value using the proper vector obtaining as BP neural network, for training and testing, finally obtain the result of statistic of classification.For the Images Classification experiment in VisTex data texturing storehouse, first from VisTex coloured image storehouse, choose at random six 512 × 512 images and test.Concrete experimental implementation step: (1) intercepts 50 samples from original image 512 × 512, picture size is 128 × 128.30 samples are for BP neural metwork training, and 20 samples are for the test of BP neural network classification; (2) use respectively multiresolution LBP texture description method, wavelet decomposition and the LBP operator that the present invention proposes to carry out texture feature extraction.(3), for MR-LBP method of the present invention, three histograms are spliced into the input vector of a LBP histogram as BP neural network; For wavelet decomposition, the variance composition characteristic vector of each component after small echo secondary decomposes is as the input value of BP neural network; For LBP operator, use
the LBP histogram of (P=8, R=1) is as the input vector of BP neural network.(4) add up respectively the correct classification rate of six texture images.Go out from the experimental results, the input vector using the proper vector that uses multiresolution LBP method of the present invention to extract as BP neural network, the accuracy of its classification compares LBP operator and wavelet decomposition has raising by a relatively large margin, particularly wavelet transformation really.This has also verified that MR-LBP method of the present invention can be good at the textural characteristics of Description Image indirectly.We are also not difficult to find simultaneously, and for the image of texture relative complex, the multiresolution LBP texture characteristic extracting method treatment effect that the present invention proposes is more outstanding.
For the classification experiments of the fresh leaf image of green tea, select the representative image of three classes, referring to Fig. 2 (a), be respectively bud one leaf, bud two leaves and a bud are leafy.Referring to Fig. 2 (b), to the fresh leaf texture image of each width tea, extract MR-LBP histogram, form 30 dimension space vectors.Specific experiment operation steps: (1) intercepts 50 samples from original image 4000 × 3000, image size is 640 × 640.30 samples are for BP neural metwork training, and 20 samples are for the test of BP neural network classification; (2) respectively with the present invention propose multiresolution LBP method, Traditional multi-scale LBP algorithm, wavelet decomposition and single invariable rotary unify LBP operator (
) carry out feature extraction.(3), for wavelet transformation, the variance composition characteristic vector of each component after small echo secondary decomposes is as the input vector of BP neural network; Unify LBP operator and Traditional multi-scale LBP algorithm for single invariable rotary, use respectively
lBP histogram and F eigenwert (with P=8, R=2,3,4 conducts
the parameter of operator, merges the proper vector of a combination of all LBP Characteristics creations, referred to as F eigenwert) as the input vector of BP neural network; The MR-LBP method proposing for the present invention, is not spliced into a multiresolution LBP histogram overlappingly by three histograms, sets it as the input vector of BP neural network; (4) add up respectively the accuracy of each method to three class Classification of Teas.Can find out from the classification of the fresh leaf of green tea, the texture extracting method of multiresolution LBP that the present invention is proposed is applied in Classification of Tea, and it is to A, B, the correct classification rate of C tri-class fresh tea leafs all, far above other three kinds of algorithms, reaches respectively 92.0%, 96.0% and 98.0%.This method for describing texture of image based on multiresolution LBP that has illustrated that the present invention proposes not only can extract the more texture information of image effectively, and is applicable to process texture relative complex and larger-size image.
Unify the texture characteristic extracting methods such as LBP operator, wavelet analysis and Traditional multi-scale LBP compares with invariable rotary, the texture characteristic extracting method (MR-LBP) of the multiresolution LBP that the present invention proposes has better texture description ability for image, and it is mainly reflected in: (1) can extract global information and the local message that image enriches more; (2) can maintain and extract the robustness of feature to rotation and the factor such as illumination and there is lower characteristic dimension; (3) for texture is relatively chaotic and larger-size complicated image treatment effect is outstanding; (4) can improve significantly the accuracy rate that fresh tea leaf is classified, classification accuracy rate is up to more than 92%.
The present invention is not limited to above-mentioned embodiment, and the present invention can expand to any new feature or any new combination disclosing in this manual, and the arbitrary new method disclosing or step or any new combination of process.
Claims (6)
1. the texture characteristic extracting method of a multiresolution LBP, it is characterized in that: (1) generates four subbands of input picture I on different frequency domains, comprising a low frequency sub-band (low-frequency approximation subgraph) and three high-frequency sub-band (high frequency details subgraph), be to obtain through one-level wavelet transform (DWT) to four described subbands, wavelet transformation can provide multiresolution analysis for image; (2) obtain after wavelet decomposition three high-frequency sub-band are averaged, generate a high-frequency average figure QUOTE
; (3) original image, low-frequency approximation subgraph and high-frequency average figure are rotated respectively to constant unified LBP and calculate, and obtain by statistics LBP histogram H (i) separately, invariable rotary is unified LBP operator, i.e. QUOTE
, be the result of invariable rotary LBP and " More General Form " combination, the dimension of described LBP histogram H (i) is 10; (4) construct 30 dimension multiresolution LBP histograms based on three LBP histograms, will after described 30 dimension multi-resolution histogram vectorizations, export.
2. the method for claim 1, is characterized in that the wavelet transform in described (1) has adopted Haar wavelet basis.
3. the method for claim 1, is characterized in that the multiresolution analysis in described (1) refers to decomposition and the expression of image being carried out to multiresolution, then under different resolution, processes respectively.
4. the method for claim 1, is characterized in that high-frequency sub-band is averaged to the signal referring on each subband correspondence position averages in described (2).
5. the method for claim 1, what it is characterized in that invariable rotary in described (3) unifies that LBP (P=8, R=1) calculates is the value between a group 0 ~ 9, the frequency formation LBP histogram occurring by adding up each value.
6. the method for claim 1, is characterized in that the multiresolution in described (4) is specially 3 kinds of different resolution.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410191280.9A CN103955691A (en) | 2014-05-08 | 2014-05-08 | Multi-resolution LBP textural feature extracting method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410191280.9A CN103955691A (en) | 2014-05-08 | 2014-05-08 | Multi-resolution LBP textural feature extracting method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN103955691A true CN103955691A (en) | 2014-07-30 |
Family
ID=51332966
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201410191280.9A Pending CN103955691A (en) | 2014-05-08 | 2014-05-08 | Multi-resolution LBP textural feature extracting method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN103955691A (en) |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105718871A (en) * | 2016-01-18 | 2016-06-29 | 成都索贝数码科技股份有限公司 | Video host identification method based on statistics |
| CN107169002A (en) * | 2017-03-31 | 2017-09-15 | 咪咕数字传媒有限公司 | A kind of personalized interface method for pushing and device recognized based on face |
| CN107729890A (en) * | 2017-11-30 | 2018-02-23 | 华北理工大学 | Face identification method based on LBP and deep learning |
| CN112712124A (en) * | 2020-12-31 | 2021-04-27 | 山东奥邦交通设施工程有限公司 | Multi-module cooperative object recognition system and method based on deep learning |
| CN113011392A (en) * | 2021-04-25 | 2021-06-22 | 吉林大学 | Pavement type identification method based on pavement image multi-texture feature fusion |
| CN113283405A (en) * | 2021-07-22 | 2021-08-20 | 第六镜科技(北京)有限公司 | Mask detection method and device, computer equipment and storage medium |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103310236A (en) * | 2013-06-27 | 2013-09-18 | 上海数据分析与处理技术研究所 | Mosaic image detection method and system based on local two-dimensional characteristics |
| CN103778434A (en) * | 2014-01-16 | 2014-05-07 | 重庆邮电大学 | Face recognition method based on multi-resolution multi-threshold local binary pattern |
-
2014
- 2014-05-08 CN CN201410191280.9A patent/CN103955691A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103310236A (en) * | 2013-06-27 | 2013-09-18 | 上海数据分析与处理技术研究所 | Mosaic image detection method and system based on local two-dimensional characteristics |
| CN103778434A (en) * | 2014-01-16 | 2014-05-07 | 重庆邮电大学 | Face recognition method based on multi-resolution multi-threshold local binary pattern |
Non-Patent Citations (3)
| Title |
|---|
| JAE YOUNG CHOI,DAE HOE KIM,YONG MAN RO: ""Combining Multiresolution Local Binary Pattern Texture Analysis and Variable Selection Strategy Applied to Computer-Aided Detection of Breast Masses on Mammograms"", 《PROCEEDINGS OF THE IEEE-EMBS INTERNATIONAL CONFERENCE ON BIOMEDICAL AND HEALTH INFORMATICS》 * |
| 宋克臣，颜云辉，陈文辉，张旭: ""一种局部二值模式方法研究与展望"", 《自动化学报》 * |
| 徐丽娟: ""基于纹理分析云的分类技术的研究"", 《中国优秀硕士学位论文全文数据库基础科学辑》 * |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105718871A (en) * | 2016-01-18 | 2016-06-29 | 成都索贝数码科技股份有限公司 | Video host identification method based on statistics |
| CN105718871B (en) * | 2016-01-18 | 2017-11-28 | 成都索贝数码科技股份有限公司 | A kind of video host's recognition methods based on statistics |
| CN107169002A (en) * | 2017-03-31 | 2017-09-15 | 咪咕数字传媒有限公司 | A kind of personalized interface method for pushing and device recognized based on face |
| CN107729890A (en) * | 2017-11-30 | 2018-02-23 | 华北理工大学 | Face identification method based on LBP and deep learning |
| CN107729890B (en) * | 2017-11-30 | 2020-10-09 | 华北理工大学 | Face recognition method based on LBP and deep learning |
| CN112712124A (en) * | 2020-12-31 | 2021-04-27 | 山东奥邦交通设施工程有限公司 | Multi-module cooperative object recognition system and method based on deep learning |
| CN113011392A (en) * | 2021-04-25 | 2021-06-22 | 吉林大学 | Pavement type identification method based on pavement image multi-texture feature fusion |
| CN113283405A (en) * | 2021-07-22 | 2021-08-20 | 第六镜科技(北京)有限公司 | Mask detection method and device, computer equipment and storage medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Tang et al. | Compressed-domain ship detection on spaceborne optical image using deep neural network and extreme learning machine | |
| CN103955691A (en) | Multi-resolution LBP textural feature extracting method | |
| CN103136766B (en) | A kind of object conspicuousness detection method based on color contrast and color distribution | |
| CN109636758B (en) | Water surface floater detection method based on space-time dynamic operator | |
| CN110415208A (en) | A kind of adaptive targets detection method and its device, equipment, storage medium | |
| CN104915676A (en) | Deep-level feature learning and watershed-based synthetic aperture radar (SAR) image classification method | |
| WO2020224221A1 (en) | Tracking method and apparatus, electronic device, and storage medium | |
| CN104574324A (en) | Denoising method for restraining spectrum characteristic of remote sensing image of ground building group | |
| CN109635789B (en) | High-resolution SAR image classification method based on intensity ratio and spatial structure feature extraction | |
| Shen et al. | Adaptive pedestrian tracking via patch-based features and spatial–temporal similarity measurement | |
| CN111612741A (en) | Accurate non-reference image quality evaluation method based on distortion recognition | |
| Zhong et al. | Robust tracking via patch-based appearance model and local background estimation | |
| Lee et al. | Multiscale morphology based illumination normalization with enhanced local textures for face recognition | |
| Gani et al. | Copy move forgery detection using DCT, PatchMatch and cellular automata | |
| Wang et al. | Image copy-move forgery detection based on dynamic threshold with dense points | |
| Gao et al. | Agricultural image target segmentation based on fuzzy set | |
| Liu et al. | Iris recognition in visible spectrum based on multi-layer analogous convolution and collaborative representation | |
| CN111160260B (en) | SAR image target detection method and system | |
| Wang et al. | Airport detection in remote sensing images based on visual attention | |
| Kovacs et al. | Orientation based building outline extraction in aerial images | |
| CN105740814B (en) | A method of determining solid waste dangerous waste storage configuration using video analysis | |
| CN106485226A (en) | A kind of video pedestrian detection method based on neutral net | |
| CN117558068B (en) | Intelligent device gesture recognition method based on multi-source data fusion | |
| Dhar et al. | Interval type-2 fuzzy set and human vision based multi-scale geometric analysis for text-graphics segmentation | |
| CN110298807A (en) | Based on the domain the NSCT infrared image enhancing method for improving Retinex and quantum flora algorithm |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20140730 |
|
| WD01 | Invention patent application deemed withdrawn after publication |