WO2014121573A1 - Procédé de détection de contours susan et système basé sur une moyenne non locale - Google Patents

Procédé de détection de contours susan et système basé sur une moyenne non locale Download PDF

Info

Publication number
WO2014121573A1
WO2014121573A1 PCT/CN2013/077101 CN2013077101W WO2014121573A1 WO 2014121573 A1 WO2014121573 A1 WO 2014121573A1 CN 2013077101 W CN2013077101 W CN 2013077101W WO 2014121573 A1 WO2014121573 A1 WO 2014121573A1
Authority
WO
WIPO (PCT)
Prior art keywords
pixel
image
edge
susan
response
Prior art date
Application number
PCT/CN2013/077101
Other languages
English (en)
Chinese (zh)
Inventor
张旭明
詹轶
吴意
丁明跃
尹周平
王瑜辉
熊有伦
Original Assignee
华中科技大学
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by 华中科技大学 filed Critical 华中科技大学
Publication of WO2014121573A1 publication Critical patent/WO2014121573A1/fr

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/13—Edge detection

Definitions

  • the invention belongs to the field of image edge detection, and particularly relates to a Susan (SUSAN) edge detection method and system based on non-local mean.
  • SUSAN Susan edge detection method and system based on non-local mean.
  • Edge detection is the basis of pattern recognition, image segmentation and image scene analysis, and is widely used in biomedical engineering and industrial automation.
  • image in actual application is inevitably polluted by noise during acquisition and transmission.
  • the introduction of noise poses a great challenge to accurate edge detection.
  • the edge detection method with excellent anti-noise performance is studied to promote its actuality. Application has important theoretical and practical significance.
  • Edge detection has been deeply studied as a hot issue in image processing and computer vision.
  • the purpose of edge detection is to identify points in the digital image where the brightness changes significantly.
  • researchers have proposed a variety of edge detection methods, such as Sobel detection method, Prewitt detection method, Canny detection method, Gabor based detection method and SUSAN edge detection method.
  • the SUSAN edge detection method has better edge detection capability than the other methods described above, and thus has received extensive attention.
  • the traditional SUSAN edge detection method uses a single pixel gray level difference to characterize pixel differences, which is susceptible to noise interference and greatly affects its noise immunity.
  • the method has superior suppression ability for noise, and can obtain edge detection results with high precision in noise pollution images.
  • Another object of the present invention is to provide a SUSAN edge detection system based on non-local mean, which has superior suppression of noise and obtains accurate edge detection results in noise-contaminated images.
  • the SUSAN edge detection method provided by the present invention specifically includes the following steps:
  • Step 1 Calculate the initial SUSAN edge response image of the image I to be detected. The process is:
  • P(P,q) is an image block centered on the pixel (p, q) in the circular template
  • P(X , y) is centered on the pixel (x, y) in the circular template.
  • Image block, image block P(p,q) and P(x,y) are the same size
  • N p is the number of pixels included in image block P(p,q)
  • * represents the convolution
  • ⁇ 2 represents the Euclidean distance
  • ⁇ is the edge detection threshold
  • ⁇ is the image An adaptive core of the same size as the block P(p,q);
  • Equation IV Calculate the final SUSAN edge response R(x', y') of the central pixel ( ⁇ ', y') of each square search window ⁇ using Equation IV, where K is the number of pixels selected in step (2.3), (i n , jj is the nth pixel selected in step (2.3);
  • R(x', y') Equation IV ⁇ w(x',y',i n ,j n ) Step 3 If the center pixel of each square search window ⁇ ( ⁇ ', y') corresponds to the final SUSAN edge The response R(x', y') is greater than the preset edge decision threshold ⁇ , then the pixel ( ⁇ ', y') is determined to be an edge pixel, otherwise the pixel (x', y') is a non-edge pixel. Further, the detected image I is also filtered before the initial SUSAN edge response image is calculated.
  • the present invention provides a SUSAN edge detection system, comprising: a first module, configured to calculate an initial SUSAN edge response image of an image to be detected I, comprising the following sub-modules:
  • a sub-module 1.1 for extracting a circular template ⁇ centered on each pixel (x, y) in the image to be detected I ;
  • the 1.2th sub-module is used to calculate the pixel difference (c, x) between each pixel ( ⁇ , q) in each circular template ⁇ and the central pixel (X, y) of the circular template by using Equation I, respectively.
  • P(p,q) and y) are image blocks centered on the pixel, q) and the circle in the circular template , the image blocks P(p, q) and P y) are the same size, and N p is the image block.
  • P(p,q) contains the number of pixels, I P ( M PI P restroom, not the image blocks P(p,q) and P y), the gray value of each pixel, * represents the convolution, ⁇ ⁇ 2 represents the Euclidean distance, the edge detection threshold is 20 ⁇ ⁇ ⁇ 35, and ⁇ is an adaptive kernel of the same size as the image block P(p, q);
  • the 1.3th sub-module is configured to calculate the initial SUSAN edge response R°(x, y) of the central pixel of each circular template according to the pixel comparison difference, that is, obtain the initial SUSAN edge response image R Q :
  • a second module configured to calculate a final SUSAN edge response image of the image to be detected I, which includes the following sub-modules:
  • the sub-module 2.2 is used to calculate the similarity between each pixel (p', q') in each square search window ⁇ and the central pixel (x', y') of the square search window by using Equation III, respectively.
  • ⁇ ( ⁇ , ⁇ are image blocks centered at (p', q') and (x', y') in the edge response image R Q , respectively, and G is a Gaussian kernel;
  • the 2.3th sub-module is used to select a portion with a large similarity from each square search window ⁇
  • the 2.4th sub-module is used to calculate the final SUSAN edge response R(x', y') of the central pixel ( ⁇ ', y') of each square search window ⁇ by using Equation IV, where K is the pixel selected by the 2.3th sub-module Number,
  • the third module is used to determine the final SUSAN edge response R(x', y corresponding to the central pixel ( ⁇ ', y') of each square search window ⁇ ') is greater than the predetermined edge decision threshold, then the pixel ( ⁇ ', y') is determined to be an edge pixel, otherwise the pixel (x', y') is a non-edge pixel. Further, a third module is further included for filtering the image to be detected I before calculating the initial SUSAN edge response image.
  • the present invention is based on a non-local mean method, which uses an image block to characterize the difference between two pixels in an image, and combines an adaptive kernel to calculate a comparison difference between pixels, thereby determining an initial SUSAN edge response of a pixel in the image, thereby overcoming
  • the existing SUSAN edge detection method based on the single pixel gray value comparison difference calculation method is susceptible to noise, can better adapt to the difference of the corresponding image structure of different pixels in the image, and lay the groundwork for accurately calculating the initial SUSAN edge response Foundation
  • the edge response can effectively reduce the isolated noise that is easily introduced by edge detection using only the initial edge response, which provides a basis for accurately calculating the final edge response.
  • FIG. 1 is a flow chart of a SUSAN edge detection method based on non-local mean according to the present invention
  • FIG. 2g is an edge image obtained by using the ISED detection method, FIG. 2h is an edge image obtained by using the ASED detection method, and FIG. 2i is an edge image detected by the method of the present invention;
  • FIG. 3 is a schematic diagram showing the comparison between the detection results of the present invention and the detection results of the other methods in Example 2, wherein FIG. 3a is a source image, FIG. 3b is an edge image obtained by TSED detection, and FIG. 3c is an edge detected by ISED. Image, Figure 3d is obtained by ASED detection The edge image, Figure 3e is the edge image detected by the method of the present invention.
  • the SUSAN edge detection method based on non-local mean of the present invention includes the following step 1: calculating an initial SUSAN edge response image R Q of the image to be detected I;
  • This step is based on the non-local mean method, which uses the image block to describe the difference between the two pixels in the image, and combines the adaptive kernel to calculate the difference between the pixels, thereby determining the initial SUSAN edge response of the pixel in the image.
  • the process is as follows:
  • P(p, q) and P(x, y) are image blocks centered on the pixels (p, q) and (x, y) in the circular template , the two image blocks are the same size, the image block The side length is smaller than the diameter of the circular template, generally 1 ⁇ 2 pixels; I P ( p ⁇ P l P consult, ⁇ , each pixel in the image block P(p, q) and P(x, y) Gray value; N p is the number of pixels included in the image block P(p, q), I P ⁇ q ⁇ P l P y) are respectively in the image blocks P(p, q) and P(x, y) n p value of gray pixels; ranges edge detection threshold T is 20 ⁇ 35; exp is an exponential function, * represents convolution, Bu 112 denotes a Euclidean distance; [Phi] of the image blocks P (p , q)
  • n(x, y) The maximum value of n(x, y) corresponding to each pixel.
  • Step 2 Calculate the final edge response image R of the image I to be detected.
  • the size of the two image blocks is the same, and the side length of the image block is smaller than the side length of the search window.
  • Rp°( p ., q ⁇ P I3 ⁇ 4 x ., y . ⁇ is the gray value of each pixel in the image block P ⁇ PP ⁇
  • G is a Gaussian kernel.
  • N is the number of pixels of the non-zero initial SUSAN edge response value in the initial SUSAN edge response image R Q ;
  • K the top K pixels with similarity in 0(x', y'). Only partial pixels with similar similarity are selected here because the smaller similarity pixels introduce noise. If K is too small, it will not achieve a good average effect. If the K value is too large, noise will be introduced to cause inaccuracy in the final edge response.
  • the recommended K value is 20 ⁇ 30, which is not limited to this, and can be adjusted according to actual conditions.
  • (i n , j n ) is the nth pixel selected in step (2.3);
  • the final SUSAN edge response with pixels is the final SUSAN edge response image R.
  • Step 3 Perform edge determination on the final SUSAN edge response image R to obtain edge pixels.
  • the specific implementation manner is as follows: If R(x', y') of the pixel (x', y') is greater than the edge determination threshold T r , then Is an edge pixel, otherwise it is a non-edge pixel.
  • the edge determination threshold T R is an empirical value, and generally takes 2 to 4.
  • the detection image I may be filtered before the initial SUSAN edge response image is calculated, and the effect is to initially smooth the image, which is beneficial to suppress noise to a certain extent and increase the accuracy of edge detection.
  • Gaussian filtering and median filtering may be adopted. And Gabor filtering, etc., Gaussian filtering is preferred.
  • the natural image is used for testing.
  • the parameters used in this example are:
  • the noise added by the natural image is Gaussian noise
  • the radius of the circular template ⁇ is 3 pixels
  • all the image blocks when calculating the pixel contrast difference c The size is 5x5, the side of the square template ⁇ is 11x11, and the similarity of the pixels is calculated.
  • the example also uses the existing Gabor-based detection method, the Canny-based detection method, the traditional SUSAN edge detection method (TSED), the improved SUSAN edge detection method (ISED), the adaptive SUSAN edge detection method (ASED) and the present.
  • the invention adopts a non-local mean SUSAN edge detection method (NLMSED) to perform edge detection on the same Gaussian noise image, and calculates a quality factor of the edge image according to the detection result (FOM: figure of merit) and a measure (P: performance measure) ) and compare.
  • 2a is a source image
  • FIG. 2c is a real edge image.
  • the edge image obtained by Gabor detection is shown in FIG.
  • FIG. 2d the edge image obtained by Canny detection is as shown in FIG.
  • Fig. 2e the edge image obtained by TSED detection is shown in Fig. 2f
  • the edge image detected by ISED is shown in Fig. 2g
  • the edge image detected by ASED is shown in Fig. 2h
  • the edge detected by NLMSED is shown.
  • the image is shown in Figure 2i.
  • the quality factor (FOM) and the performance measure (P) were used as the measure for evaluating the edge detection quality, and the parameters were measured as shown in Table 1.
  • the CT image is used for testing.
  • the parameters used in this example are:
  • the original noise in the image can be considered as Gaussian noise, the radius of the circular template ⁇ is 4 pixels, and all the image blocks when calculating the pixel contrast difference c
  • the size is 5x5, the side length of the square search window ⁇ is 11x11, and the similarity of the pixels is calculated.
  • This example uses the existing Gabor-based detection method, the Canny-based detection method, the traditional SUSAN edge detection method (TSED), the improved SUSAN edge detection method (ISED), the adaptive SUSAN edge detection method (ASED), and the present invention.
  • the SUSAN edge detection method (NLMSED) based on non-local mean performs edge detection on the same Gaussian noise image.
  • Figure 3a is the source noise image
  • the edge image obtained by TSED is shown in Figure 3b
  • the edge image detected by ISED is shown in Figure 3c
  • the edge image detected by ASED is shown in Figure 3d.
  • the edge image detected by NLMSED is shown in Figure 3e.
  • the NLMSED method of the present invention obtains sharper edges, better continuity, less residual noise, and better noise resistance than other comparison methods.

Landscapes

  • Engineering & Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Image Analysis (AREA)

Abstract

L'invention concerne un procédé de détection de contours SUSAN basé sur une moyenne non locale, ledit procédé comprenant : une étape consistant à calculer une image de réponse de contour SUSAN initiale : pour chaque pixel (x, y) dans l'image, prendre un modèle circulaire Ψ avec celui-ci comme centre, présenter la différence de comparaison des pixels de chaque pixel dans le Ψ au moyen d'un bloc d'image, et calculer la réponse de contour SUSAN initiale R0 des pixels (x, y) conformément à la différence de comparaison des pixels; une étape consistant à calculer une image de réponse de contour SUSAN finale : pour chaque pixel (x', y') dans le R0, rechercher un voisinage présentant une structure similaire à celui-ci, et calculer la moyenne de similitude pondérée de la réponse de contour SUSAN initiale pour obtenir la réponse de contour SUSAN finale R des pixels (x', y'); et une étape consistant à évaluer un pixel de contour : si le R (x', y') correspondant au pixel (x', y') est supérieur à un seuil, le pixel est un pixel de contour; dans le cas contraire, le pixel n'est pas un pixel de contour. L'invention concerne également un système permettant de réaliser le procédé. En adoptant le procédé de détection de contour SUSAN, la moyenne non locale est appliquée à la détection de contours SUSAN afin d'améliorer sensiblement l'immunité au bruit pour obtenir une précision plus élevée de la détection de contours.
PCT/CN2013/077101 2013-02-06 2013-06-09 Procédé de détection de contours susan et système basé sur une moyenne non locale WO2014121573A1 (fr)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201310047546.8 2013-02-06
CN201310047546.8A CN103150725B (zh) 2013-02-06 2013-02-06 基于非局部均值的susan边缘检测方法及系统

Publications (1)

Publication Number Publication Date
WO2014121573A1 true WO2014121573A1 (fr) 2014-08-14

Family

ID=48548778

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2013/077101 WO2014121573A1 (fr) 2013-02-06 2013-06-09 Procédé de détection de contours susan et système basé sur une moyenne non locale

Country Status (2)

Country Link
CN (1) CN103150725B (fr)
WO (1) WO2014121573A1 (fr)

Families Citing this family (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103700076B (zh) * 2013-12-26 2016-09-14 辽宁师范大学 基于非局部均值法的视频图像快速去噪方法
CN104200442B (zh) * 2014-09-19 2017-11-21 西安电子科技大学 基于改进的canny边缘检测的非局部均值MRI图像去噪方法
CN108022233A (zh) * 2016-10-28 2018-05-11 沈阳高精数控智能技术股份有限公司 一种基于改进型Canny算子的工件边缘提取方法
CN106651807B (zh) * 2016-12-29 2020-03-10 上海天马有机发光显示技术有限公司 一种图像处理系统、显示设备及图像处理方法
CN107833206B (zh) * 2017-10-24 2021-07-06 武汉大学 一种复杂背景下电力线精确提取方法
CN109949298B (zh) * 2019-03-22 2022-04-29 西南交通大学 一种基于聚类学习的图像分割质量评价方法

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN1686051A (zh) * 2005-05-08 2005-10-26 上海交通大学 基于vpf和改进的susan的眼角和瞳孔的定位方法
US20080253678A1 (en) * 2007-04-10 2008-10-16 Arcsoft, Inc. Denoise method on image pyramid
US20110102852A1 (en) * 2009-10-30 2011-05-05 Xerox Corporation Susan-based corner sharpening
CN102663692A (zh) * 2012-03-28 2012-09-12 汕头大学 医学超声图像自适应susan扩散去噪方法

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102339464A (zh) * 2011-09-01 2012-02-01 上海大学 线搜索式角点检测方法

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN1686051A (zh) * 2005-05-08 2005-10-26 上海交通大学 基于vpf和改进的susan的眼角和瞳孔的定位方法
US20080253678A1 (en) * 2007-04-10 2008-10-16 Arcsoft, Inc. Denoise method on image pyramid
US20110102852A1 (en) * 2009-10-30 2011-05-05 Xerox Corporation Susan-based corner sharpening
CN102663692A (zh) * 2012-03-28 2012-09-12 汕头大学 医学超声图像自适应susan扩散去噪方法

Also Published As

Publication number Publication date
CN103150725B (zh) 2015-09-23
CN103150725A (zh) 2013-06-12

Similar Documents

Publication Publication Date Title
WO2014121573A1 (fr) Procédé de détection de contours susan et système basé sur une moyenne non locale
CN107301661B (zh) 基于边缘点特征的高分辨率遥感图像配准方法
CN102098440B (zh) 针对摄像机抖动下运动目标检测的电子稳像方法及系统
TWI345916B (en) Motion estimation method
CN102156996B (zh) 一种图像边缘检测的方法
CN104217416B (zh) 灰度图像处理方法及其装置
CN108932699B (zh) 基于变换域的三维匹配调和滤波图像去噪方法
CN110070574B (zh) 一种基于改进PSMNet的双目视觉立体匹配方法
CN109214380A (zh) 车牌倾斜校正方法
CN113837198B (zh) 一种基于三维块匹配的改进自适应阈值Canny边缘检测方法
CN104463911A (zh) 基于复杂背景估计的红外运动小目标检测方法
RU2012152935A (ru) Устройство распознавания жестов, способ распознавания жестов и программа
CN104778707A (zh) 一种改进广义Hough变换的电解电容检测方法
CN108647597B (zh) 一种手腕识别方法、手势识别方法、装置和电子设备
CN104036465A (zh) 基于边缘检测的自适应非局部均值去噪方法
CN107360435B (zh) 块效应检测方法、块噪声滤除方法及装置
CN105389793B (zh) 一种自动识别物体表面裂缝走向及宽度的方法
CN110992288B (zh) 一种用于矿山井筒环境下的视频图像盲去噪方法
CN107993219A (zh) 一种基于机器视觉的船板表面缺陷检测方法
CN111429372A (zh) 一种增强低对比度图像边缘检测效果的方法
CN110245600A (zh) 自适应起始快速笔画宽度无人机道路检测方法
CN113436216B (zh) 基于Canny算子的电气设备红外图像边缘检测方法
Xue et al. Edge detection combining wavelet transform and canny operator based on fusion rules
CN105389820A (zh) 一种基于倒谱的红外图像清晰度评价方法
WO2014205787A1 (fr) Procédé de détection de véhicule basé sur un modèle d'image hybride

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 13874506

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 13874506

Country of ref document: EP

Kind code of ref document: A1