CN104573713A - Mutual inductor infrared image recognition method based on image textual features - Google Patents

Abstract

The invention provides a mutual inductor infrared image recognition method based on image textual features. The method includes the steps of conducting graying and denoising on an image, and conducting the X-direction Canny detection on the image, wherein the textural features of the left edge an the right edge of a sleeve are exposed after graying, denoising and detection are conducted; fitting left feature points into a straight line and fitting right feature points into a straight line after the methods of the normalized correlation coefficient maximum value matching and K-means clustering are conducted, and determining the highest edge points and the lowest edge points on the straight lines so as to determine the image area of a mutual inductor. By means of the method, the defect that a traditional image recognition method is not enough in accuracy is overcome, the background interference prevention capacity is high, the specific area of the mutual inductor in the infrared thermal image can be rapidly recognized, operating personnel are helped to detect the position coordinates of the mutual inductor, the automation degree of the state monitoring of a power system is improved, and the very high practical value is achieved.

Description

A kind of mutual inductor Infrared image recognition based on image texture characteristic

Technical field

The invention belongs to pattern-recognition and technical field of computer vision, particularly relate to a kind of mutual inductor Infrared image recognition based on image texture characteristic.

Background technology

Video monitoring system has been installed in power plant, transformer station by some Utilities Electric Co. at present, can realize monitoring field apparatus, controlling the functions such as remote camera action.But these video monitoring systems only have video monitoring function, there is no video image automatic identification function.For giving full play to the function of video monitoring system, judge the reason of the on-the-spot alarm that has an accident more accurately, remote digital video monitoring and digital image acquisition system should be adopted, to realize the image recognition of equipment alarm, new means are provided, for crash analysis provides reliable foundation for accident detects.

Domestic and international at present still planless discussion is identified automatically to the mutual inductor based on infrared image textural characteristics." remote image supervision of sub station system " article of Tang Huiming, Zhang Jian proposes a kind of electrical equipment image recognition methods based on image Hu not bending moment, the method calculates the Hu not bending moment of this bianry image, constitute the proper vector of power equipment, devise BP neural network classifier and do Classification and Identification, can be used for the Fault Identification realizing power equipment infrared image in electric system in conjunction with temperature information.But the method only can identify in image the electrical equipment that there is certain type, and fail to realize the accurate location of electrical equipment, therefore have certain limitation.Chen Junyou, Gionee army, Duan Shaohui, Yao Senjing, Zhao Ling survey article " the infrared image power equipment identification based on Hu not bending moment " and propose one and utilize infrared image and neural network weight Weigh Direct Determination to carry out on-the-spot zero resistance insulator to know method for distinguishing, but it is only limitted to the identification field of insulator, do not relate to mutual inductor identification, method simultaneously is herein only applicable to the few infrared image of background interference, and the actual conditions of this and substation field are not inconsistent, on-the-spot bad adaptability.In addition, most of achievement in research mainly to be extracted target based on the gray difference between mutual inductor and background and is identified.But due to background interference, the accuracy of identification is not high.For this reason, be necessary to develop a kind of Infrared image recognition for mutual inductor texture feature extraction to improve the accuracy and specific aim that identify, but existing document or pay close attention to integral image discriminator and can not the accurate location of realize target, or pay close attention to the location of target device, identify that the research of accuracy is but comparatively limited to improving.

The image recognition of extracting based on image texture characteristic is now applied in the multiple field of pattern-recognition, comprises mechanical component identification, Aero-Space, agricultural engineering etc.Image-recognizing method for mutual inductor self texture feature extraction is that the accurate location of mutual inductor in infrared image and identification provide the foundation.

Summary of the invention

The present invention proposes a kind of mutual inductor Infrared image recognition based on image texture characteristic, object is to improve traditional images recognition methods accuracy.

The present invention is intended to overcome the deficiencies in the prior art, take into full account the feature of substation field background complexity, take into full account the image texture characteristic of mutual inductor edge " waveform ", normalized correlation coefficient maximal value coupling and K-means clustering method are incorporated, a kind of mutual inductor Infrared image recognition based on image texture characteristic is proposed, the method recognition result can react mutual inductor position coordinate data in the picture objectively, for follow-up temperature and Fault Identification lay the foundation, promote the automaticity that electric network fault differentiates, for achieving the above object, the technical solution used in the present invention is: a kind of mutual inductor Infrared image recognition based on image texture characteristic, comprise the steps:

(1) Image semantic classification, comprises mutual inductor image gray processing and denoising, and carries out the Canny detection of X-direction to image, extracts the feature texture of mutual inductor;

(2) adopt the method for normalized correlation coefficient maximal value to mate pretreatment image and left and right edges characteristic feature image, the texture characteristic points realizing image is extracted;

(3) K-means clustering processing is carried out to unique point, and reject outlier;

(4) according to the cluster result of left and right edges, cluster centre is carried out to the fitting a straight line of least square method, determine the left and right edges of sleeve pipe;

(5) on the edge line of matching, find the highest and minimum point of ordinate, determine the lower edges of sleeve pipe, finally obtain the mutual inductor region identified.

Further, described step (1) specifically comprises the steps:

A) method of image gray processing process is as follows:

Gray (i, j)=0.299 × r (i, j)+0.587 × g (i, j)+0.114 × b (i, j) formula (1)

Wherein, gray (i, j) represents (i, j) gray-scale value put, r (i, j) represents that (i, j) puts the R value component of rgb value, g (i, j) represent that (i, j) puts the G value component of rgb value, b (i, j) represent that (i, j) puts the B value component of rgb value;

B) image denoising adopts the method for Butterworth low pass ripple, and the transport function of a n rank Butterworth filter is:

$H(u,v)=\frac{1}{1+{\left[\frac{D(u,v)}{D_0}\right]}^{2n}}$ Formula (2)

Wherein, D ₀it is cutoff frequency; N is exponent number, gets positive integer, thus the shape of controlling curve, because Butterworth filter transition curve is comparatively level and smooth, therefore the fuzzy of image will reduce;

C) the x direction gradient value of image is after pretreatment:

Gray _x(i, j)=gray (i+1, j)-gray (i-1, j) formula (3)

In formula (3), by the center pixel of neighborhood compared with the interpolation result of two gradient magnitudes along gradient direction, if the amplitude of the centre of neighbourhood is large unlike two interpolation results of gradient direction, be then 0 by the edge flag position assignment that gray (i, j) is corresponding.

Further, described step (2) specifically comprises the steps:

A) calculate normalized correlation coefficient, computing method are as follows:

$\mathrm{RN}(x,y)=\frac{\underset{p=0}{\overset{P-1}{\mathrm{\Σ}}}\underset{q=0}{\overset{Q-1}{\mathrm{\Σ}}}(t(p,q)-\stackrel{\‾}{t})(f(x+p,y+q)-\stackrel{\‾}{f})}{\sqrt{\underset{p=0}{\overset{P-1}{\mathrm{\Σ}}}\underset{q=0}{\overset{K-1}{\mathrm{\Σ}}}{[f(x+p,y+q)-\stackrel{\‾}{f}]}^2}\sqrt{\underset{p=0}{\overset{P-1}{\mathrm{\Σ}}}\underset{q=0}{\overset{Q-1}{\mathrm{\Σ}}}{\left[(t(p,q)-\stackrel{\‾}{t})\right]}^2}}$ (formula 4)

Wherein

$\begin{array}{cc}\stackrel{\‾}{t}=\frac{\underset{p=0}{\overset{P-1}{\mathrm{\Σ}}}\underset{q=0}{\overset{Q-1}{\mathrm{\Σ}}}t(p,q)}{P\×Q}& \stackrel{\‾}{f}=\frac{\underset{p=0}{\overset{P-1}{\mathrm{\Σ}}}\underset{q=0}{\overset{Q-1}{\mathrm{\Σ}}}f(x+p,y+q)}{P\×Q}\end{array}$ (formula 5)

In formula, the original image that f (x, y) is M × N, namely through pretreated X-direction Canny detected image, t (p, q) for the template image of P × Q (P≤M, Q≤N), RN (x, y) be normalized correlation coefficient;

B) point of filtering RN (x, y) <0.7, remaining point is considered to a left side (right side) edge feature of mutual inductor.

Further, described step (3) specifically comprises the steps:

A) method of K-means cluster is as follows:

$E=\underset{i=l}{\overset{K^{\&prime;}}{\mathrm{\&Sigma;}}}\underset{x\&Element;C_i}{\mathrm{\&Sigma;}}|X-{\stackrel{\&OverBar;}{X}}_i|$ (formula 6)

K' is clusters number, selected 14 cluster centres, the first initial cluster center of Stochastic choice 14 classes, to any one object, calculate the distance at these 14 centers, and this object is grouped in the class at the maximum place, center of similarity, utilize averaging method to upgrade the central value of k class, to all 14 cluster centres, if after iterative algorithm upgrades, measure function shown in formula (6) is restrained or is reached maximum iteration time, then iteration terminates, otherwise continues iteration;

B) set the coordinate figure of 14 cluster centres after K-means cluster as (i _w, j _w), wherein w=1,2...14, calculates the Euclidean distance of 1 ~ 14 cluster centre point to other 13 cluster centre points respectively, and form 14 × 14 rank Euclidean distance matrixes, Euclidean distance matrix by rows is added and forms Euclidean distance and matrix dis (u, 1) u=1,2...14, again by Euclidean distance and from small to large arrangement formed dis'(u, 1) u=1,2...14 matrix, when time, assert that u point is outlier.

Further, described step (4) specifically comprises the steps:

A) to reject the cluster centre point set coordinate of outlier for sample, adopt least square method to carry out fitting a straight line, concrete formula is as follows:

formula (7)

formula (8)

formula (9)

Wherein, o represents the cluster centre number after 14 cluster centre points rejecting outlier, x _i, y _ibe respectively the transverse and longitudinal coordinate figure of cluster centre point, be respectively the mean value of cluster centre point transverse and longitudinal coordinate, (9) formula is the straight-line equation of matching;

B) a mutual inductor left side (right side) Edge Feature Points adding RN (x, y) <0.7 carrys out the accuracy of accentuated edges identification, carries out edge revision.Assuming that φ { ii _ss, jj _ssbe Edge Feature Points set, ss is the numbering of Edge Feature Points, ii _ss, jj _ssbe respectively its transverse and longitudinal coordinate figure, if

formula (10)

Namely (ii is thought _ss, jj _ss) be actual Edge Feature Points, (10) formula is a little to the range formula of straight line, confirms through experiment, when selected Edge Feature Points distance feature air line distance is within 3 pixels, can includes maximum Edge Feature Points, cause least error simultaneously.If do not meet (10) formula, then think the Edge Feature Points peeled off, final composition actual edge unique point set Φ { ii _ss, jj _ss, the coordinate of this set mid point is brought (7) into ~ (9) formula, finally obtain the edge line equation revised.

Further, described step (5) specifically comprises the steps:

Determine the highs and lows ordinate of mutual inductor left and right edges actual edge unique point, thus determine the straight line of the lower edges of sleeve pipe, do lower edges two straight lines and with left and right edges matching correction straight line intersection, middle region is the final mutual inductor region identified.

The advantage that the present invention has and good effect are: the mutual inductor Infrared image recognition based on image texture characteristic proposed by the present invention, the feature texture of mutual inductor can be extracted according to " waveform " feature of mutual inductor left and right edges, after K-means cluster and outlier are rejected, fit to left and right edges straight line, and the feature correction edge line of more multiple edge information is contained according to feature texture, the present invention can identify the concrete region of mutual inductor in infrared thermal imagery rapidly, operations staff is helped to detect the position coordinates of mutual inductor, improve the automaticity of POWER SYSTEM STATE monitoring, there is very strong practical value.

Accompanying drawing explanation

Fig. 1 is the schematic diagram of the mutual inductor infrared image of substation field shooting;

Fig. 2 is the schematic diagram of mutual inductor infrared image image after pretreatment;

Fig. 3 is the schematic diagram of the template image of mutual inductor left and right edges feature;

Fig. 4 is the schematic diagram of the mutual inductor left and right edges unique point image after normalized correlation coefficient maximal value coupling;

Fig. 5 is the schematic diagram of the result schematic diagram of mutual inductor left hand edge unique point after K-means cluster;

Fig. 6 is the schematic diagram of the result schematic diagram of mutual inductor right hand edge unique point after K-means cluster;

Fig. 7 is the schematic diagram of the Euclidean distance matrix schematic diagram between cluster centre point;

Fig. 8 is the schematic diagram of the schematic diagram of left and right edges fitting a straight line;

Fig. 9 is the schematic diagram of the fitted straight lines of edges schematic diagram revised after adding Edge Feature Points;

Figure 10 is the schematic diagram of the position view of mutual inductor in infrared image.

Embodiment

Below by way of the mutual inductor Infrared image recognition based on image texture characteristic of the present invention being applied in the mutual inductor infrared image identification of Tianjin Baodi Utilities Electric Co. collection.

In this recognition system, comprise Baodi power supply administration and have transformer station's mutual inductor equipment under its command, cover vegetable garden great Kou Tun, the new opening in area just outside a city gate, east old mouth Zhou Liang, with village, peaceful village Deng Di transformer station mutual inductor equipment, comprise the mutual inductor infrared image of 320 shootings altogether, the mutual inductor size in these infrared images and angle all variant.The present invention proposes mutual inductor recognition methods, and for identifying the particular location of mutual inductor in these infrared images, and extract from image so that follow-up temperature and Fault Identification, concrete steps are as follows:

Step 1: pre-service is carried out to the mutual inductor infrared image in Fig. 1, first remove below numerical information and right than vitta region, then image gray processing and denoising is carried out, finally image is carried out to the Canny detection of X-direction, extract the feature texture of mutual inductor, complete the pre-service of image.

The pre-service of image specifically comprises the steps:

A) method of image gray processing process is as follows:

Gray (i, j)=0.299 × r (i, j)+0.587 × g (i, j)+0.114 × b (i, j) formula (1)

Wherein, gray (i, j) represents (i, j) gray-scale value put, r (i, j) represents that (i, j) puts the R value component of rgb value, g (i, j) represent that (i, j) puts the G value component of rgb value, b (i, j) represent that (i, j) puts the B value component of rgb value.

B) image denoising adopts the method for Butterworth low pass ripple, and the transport function of a n rank Butterworth filter is:

$H(u,v)=\frac{1}{1+{\left[\frac{D(u,v)}{D_0}\right]}^{2n}}$ Formula (2)

Wherein, D ₀it is cutoff frequency; N is exponent number, gets positive integer, thus the shape of controlling curve, because Butterworth filter transition curve is comparatively level and smooth, therefore the fuzzy of image will reduce.

C) the x direction gradient value of image is after pretreatment:

Gray _x(i, j)=gray (i+1, j)-gray (i-1, j) formula (3)

In formula (3), by the center pixel of neighborhood compared with the interpolation result of two gradient magnitudes along gradient direction, if the amplitude of the centre of neighbourhood is large unlike two interpolation results of gradient direction, be then 0 by the edge flag position assignment that gray (i, j) is corresponding.

Final result as shown in Figure 2.

Step 2: adopt the method for normalized correlation coefficient maximal value to mate pretreatment image and left and right edges characteristic feature image, the texture characteristic points realizing image is extracted.Fig. 3 is the mutual inductor left and right edges feature templates image of standard, take Fig. 2 as original image,

If the original image that f (x, y) is M × N, namely through pretreated X-direction Canny detected image, t (p, q) is the template image of P × Q (P≤M, Q≤N), then coefficient R N (x, y) can be expressed as:

$\frac{\underset{p=0}{\overset{P-1}{\mathrm{\&Sigma;}}}\underset{q=0}{\overset{Q-1}{\mathrm{\&Sigma;}}}(t(p,q)-\stackrel{\&OverBar;}{t})(f(x+p,y+q)-\stackrel{\&OverBar;}{f})}{\sqrt{\underset{p=0}{\overset{P-1}{\mathrm{\&Sigma;}}}\underset{q=0}{\overset{K-1}{\mathrm{\&Sigma;}}}{[f(x+p,y+q)-\stackrel{\&OverBar;}{f}]}^2}\sqrt{\underset{p=0}{\overset{P-1}{\mathrm{\&Sigma;}}}\underset{q=0}{\overset{Q-1}{\mathrm{\&Sigma;}}}{\left[(t(p,q)-\stackrel{\&OverBar;}{t})\right]}^2}}$ Formula (4)

Wherein

$\begin{array}{cc}\stackrel{\&OverBar;}{t}=\frac{\underset{p=0}{\overset{P-1}{\mathrm{\&Sigma;}}}\underset{q=0}{\overset{Q-1}{\mathrm{\&Sigma;}}}t(p,q)}{P\&times;Q}& \stackrel{\&OverBar;}{f}=\frac{\underset{p=0}{\overset{P-1}{\mathrm{\&Sigma;}}}\underset{q=0}{\overset{Q-1}{\mathrm{\&Sigma;}}}f(x+p,y+q)}{P\&times;Q}\end{array}$ Formula (5)

After formula (4) ~ (5) calculate, remove the point that normalized correlation coefficient is less than 0.7, result as shown in Figure 4.

Step 3: respectively K-means cluster is carried out to all left hand edge unique points and right hand edge unique point, gathering respectively is 14 classes, and form 28 cluster centre points altogether, method is as follows:

$E=\underset{i=l}{\overset{K^{\&prime;}}{\mathrm{\&Sigma;}}}\underset{x\&Element;C_i}{\mathrm{\&Sigma;}}|X-{\stackrel{\&OverBar;}{X}}_i|$ Formula (6)

Cluster result is as shown in Fig. 5 ~ 6.Then compute euclidian distances matrix, if the coordinate figure of 14 cluster centres after K-means cluster is (i _w, j _w), wherein w=1,2...14, calculates the Euclidean distance of 1 ~ 14 cluster centre point to other 13 cluster centre points respectively, and form 14 × 14 rank Euclidean distance matrixes, Euclidean distance matrix by rows is added and forms Euclidean distance and matrix dis (u, 1) u=1,2...14, again by Euclidean distance and from small to large arrangement formed dis'(u, 1) u=1,2...14 matrix, when time, assert that u point is outlier.Result as shown in Figure 7, has three outlier in this mutual inductor left hand edge unique point, after outlier being rejected, forms final cluster centre set.

Step 4: carry out least square line matching to left hand edge cluster centre meeting point and right hand edge cluster centre meeting point respectively, concrete formula is as follows:

formula (7)

formula (8)

formula (9)

Wherein, o represents the cluster centre number after 14 cluster centre points rejecting outlier, x _i, y _ibe respectively the transverse and longitudinal coordinate figure of cluster centre point, x, y is respectively the mean value of cluster centre point transverse and longitudinal coordinate, and (9) formula is the straight-line equation of matching, and result as shown in Figure 8, for improving the accuracy of limb recognition, a mutual inductor left side (right side) Edge Feature Points adding RN (x, y) <0.7 carrys out the accuracy of accentuated edges identification, again because the straight line of matching is accurately approximate, therefore edge correction can be carried out, assuming that φ { ii by based on _ss, jj _ssbe Edge Feature Points set, ss is the numbering of Edge Feature Points, ii _ss, jj _ssbe respectively its transverse and longitudinal coordinate figure, if

formula (10)

Namely (ii is thought _ss, jj _ss) be actual Edge Feature Points, if do not meet (10) formula, then think the Edge Feature Points peeled off, final composition actual edge unique point set Φ { ii _ss, jj _ss, the coordinate of this set mid point is brought (7) into ~ (9) formula, finally obtain the edge line equation revised, revise fitting a straight line result by formula (10), result is as shown in Figure 9.

Step 5: assuming that the highs and lows ordinate of mutual inductor left hand edge actual edge unique point divides left _max, left _min, the ordinate of the highs and lows of right hand edge actual edge unique point is respectively right _max, right _min.

To sum up, coboundary peak ordinate max=max{left _max, right _max, lower limb minimum point ordinate min=min{left _min, right _min, do y=max and y=min two straight lines respectively and with left and right edges matching correction straight line intersection, middle region is the final mutual inductor region identified, mutual inductor position view is as shown in Figure 10.

Above embodiments of the invention have been described in detail, but described content being only preferred embodiment of the present invention, can not being considered to for limiting practical range of the present invention.All equalizations done according to the scope of the invention change and improve, and all should still belong within this patent covering scope.

Claims (6)

1., based on a mutual inductor Infrared image recognition for image texture characteristic, comprise the steps:

(1) Image semantic classification, comprises mutual inductor image gray processing and denoising, and carries out the Canny detection of X-direction to image, extracts the feature texture of mutual inductor;

(2) adopt the method for normalized correlation coefficient maximal value to mate pretreatment image and left and right edges characteristic feature image, the texture characteristic points realizing image is extracted;

(3) K-means clustering processing is carried out to unique point, and reject outlier;

(4) according to the cluster result of left and right edges, cluster centre is carried out to the fitting a straight line of least square method, determine the left and right edges of sleeve pipe;

(5) on the edge line of matching, find the highest and minimum point of ordinate, determine the lower edges of sleeve pipe, finally obtain the mutual inductor region identified.

2. a kind of mutual inductor Infrared image recognition based on image texture characteristic according to claim 1, is characterized in that: described step (1) specifically comprises the steps:

A) method of image gray processing process is as follows:

Gray (i, j)=0.299 × r (i, j)+0.587 × g (i, j)+0.114 × b (i, j) formula (1)

Wherein, gray (i, j) represents (i, j) gray-scale value put, r (i, j) represents that (i, j) puts the R value component of rgb value, g (i, j) represent that (i, j) puts the G value component of rgb value, b (i, j) represent that (i, j) puts the B value component of rgb value;

B) image denoising adopts the method for Butterworth low pass ripple, and the transport function of a n rank Butterworth filter is:

$H(u,v)=\frac{1}{1+{\left[\frac{D(u,v)}{D_0}\right]}^{2n}}$ Formula (2)

Wherein, D ₀it is cutoff frequency; N is exponent number, gets positive integer, thus the shape of controlling curve, because Butterworth filter transition curve is comparatively level and smooth, therefore the fuzzy of image will reduce;

C) the x direction gradient value of image is after pretreatment:

Gray _x(i, j)=gray (i+1, j)-gray (i-1, j) formula (3)

In formula (3), by the center pixel of neighborhood compared with the interpolation result of two gradient magnitudes along gradient direction, if the amplitude of the centre of neighbourhood is large unlike two interpolation results of gradient direction, be then 0 by the edge flag position assignment that gray (i, j) is corresponding.

3. a kind of mutual inductor Infrared image recognition based on image texture characteristic according to claim 1, is characterized in that: described step (2) specifically comprises the steps:

A) calculate normalized correlation coefficient, computing method are as follows:

$\mathrm{RN}(x,y)=\frac{\underset{p=0}{\overset{P-1}{\mathrm{\&Sigma;}}}\underset{q=0}{\overset{Q-1}{\mathrm{\&Sigma;}}}(t(p,q)-\stackrel{\&OverBar;}{t})(f(x+p,y+q)-\stackrel{\&OverBar;}{f})}{\sqrt{\underset{p=0}{\overset{P-1}{\mathrm{\&Sigma;}}}\underset{q=0}{\overset{K-1}{\mathrm{\&Sigma;}}}{\left[(f(x+p,y+q)-\stackrel{\&OverBar;}{f})\right]}^2}\sqrt{\underset{p=0}{\overset{P-1}{\mathrm{\&Sigma;}}}\underset{q=0}{\overset{Q-1}{\mathrm{\&Sigma;}}}{\left[(t(p,q)-\stackrel{\&OverBar;}{t})\right]}^2}}$ (formula 4)

Wherein

$\begin{array}{cc}\stackrel{\&OverBar;}{t}=\frac{\underset{p=0}{\overset{P-1}{\mathrm{\&Sigma;}}}\underset{q=0}{\overset{Q-1}{\mathrm{\&Sigma;}}}t(p,q)}{P\&times;Q}& \stackrel{\&OverBar;}{f}=\frac{\underset{p=0}{\overset{P-1}{\mathrm{\&Sigma;}}}\underset{q=0}{\overset{Q-1}{\mathrm{\&Sigma;}}}f(x+p,y+q)}{P\&times;Q}\end{array}$ (formula 5)

In formula, the original image that f (x, y) is M × N, namely through pretreated X-direction Canny detected image, t (p, q) for the template image of P × Q (P≤M, Q≤N), RN (x, y) be normalized correlation coefficient;

B) point of filtering RN (x, y) <0.7, a left side (right side) edge that remaining point is considered to mutual inductor is special

Levy.

4. a kind of mutual inductor Infrared image recognition based on image texture characteristic according to claim 1, is characterized in that: described step (3) specifically comprises the steps:

A) method of K-means cluster is as follows:

$E=\underset{i=1}{\overset{K^{\&prime;}}{\mathrm{\&Sigma;}}}\underset{x\&Element;C_i}{\mathrm{\&Sigma;}}|X-\stackrel{\&OverBar;}{X_i}|$ (formula 6)

K' is clusters number, selected 14 cluster centres, the first initial cluster center of Stochastic choice 14 classes, to any one object, calculate the distance at these 14 centers, and this object is grouped in the class at the maximum place, center of similarity, utilize averaging method to upgrade the central value of k class, to all 14 cluster centres, if after iterative algorithm upgrades, measure function shown in formula (6) is restrained or is reached maximum iteration time, then iteration terminates, otherwise continues iteration;

B) set the coordinate figure of 14 cluster centres after K-means cluster as (i _w, j _w), wherein w=1,2...14, calculates the Euclidean distance of 1 ~ 14 cluster centre point to other 13 cluster centre points respectively, and form 14 × 14 rank Euclidean distance matrixes, Euclidean distance matrix by rows is added and forms Euclidean distance and matrix dis (u, 1) u=1,2...14, again by Euclidean distance and from small to large arrangement formed dis'(u, 1) u=1,2...14 matrix, when time, assert that u point is outlier.

5. a kind of mutual inductor Infrared image recognition based on image texture characteristic according to claim 1, is characterized in that: described step (4) specifically comprises the steps:

A) to reject the cluster centre point set coordinate of outlier for sample, adopt least square method to carry out fitting a straight line, concrete formula is as follows:

formula (7)

formula (8)

formula (9)

Wherein, o represents the cluster centre number after 14 cluster centre points rejecting outlier, x _i, y _ibe respectively the transverse and longitudinal coordinate figure of cluster centre point, be respectively the mean value of cluster centre point transverse and longitudinal coordinate, (9) formula is the straight-line equation of matching;

B) a mutual inductor left side (right side) Edge Feature Points adding RN (x, y) <0.7 carrys out the accuracy of accentuated edges identification, carries out edge revision.Assuming that φ { ii _ss, jj _ssbe Edge Feature Points set, ss is the numbering of Edge Feature Points, ii _ss, jj _ssbe respectively its transverse and longitudinal coordinate figure, if

Namely (ii is thought _ss, jj _ss) be actual Edge Feature Points, (10) formula is a little to the range formula of straight line, confirms through experiment, when selected Edge Feature Points distance feature air line distance is within 3 pixels, can includes maximum Edge Feature Points, cause least error simultaneously.If do not meet (10) formula, then think the Edge Feature Points peeled off, final composition actual edge unique point set Φ { ii _ss, jj _ss, the coordinate of this set mid point is brought (7) into ~ (9) formula, finally obtain the edge line equation revised.

6. a kind of mutual inductor Infrared image recognition based on image texture characteristic according to claim 1, is characterized in that: described step (5) specifically comprises the steps:

Determine the highs and lows ordinate of mutual inductor left and right edges actual edge unique point, thus determine the straight line of the lower edges of sleeve pipe, do lower edges two straight lines and with left and right edges matching correction straight line intersection, middle region is the final mutual inductor region identified.