CN104200215A - Method for identifying dust and pocking marks on surface of big-caliber optical element - Google Patents

Abstract

The invention discloses a method for identifying dust and pocking marks on a surface of a big-caliber optical element. The method comprises the following steps of: utilizing an image acquisition device to acquire surface damage images of the big-caliber optical element under the condition of a light field to serve as sample images, and marking the damage type corresponding to each sample image; extracting characteristic vectors of the acquired sample images; performing pre-treatment on the obtained characteristic vectors; regarding the characteristic vectors after the pre-treatment as input variables of a mode identification classifier, and regarding corresponding damage type labels as output variables of a mode identification classifier so as to obtain a mapping relationship between the input variables and the output variables; identifying the damage types of the tested samples based on the mapping relationship, and outputting the identification result. The mode identification method is adopted to start from the texture characteristics of images in the light field, so that the automatic judgment problem of the dust and the pocking marks on the surface of the big-caliber optical element is solved.

Description

A kind of optical elements of large caliber surface dirt and pit recognition methods

Technical field

The present invention relates to mode identification technology, especially a kind of optical elements of large caliber surface dirt and pit recognition methods.

Background technology

Optical elements of large caliber is widely used in the optical-mechanical system of the high-light-energy current density taking modern laser as representative, the low-light level imaging field taking infrared night vision optical system as representative and the field of semiconductor processing and manufacturing taking crystal column surface processing detection as representative.Optical elements of large caliber can produce all kinds of defects in the processes such as generation, processing, cleaning and transport, such as pit, cut, bound edge, deliquescence class, mildew, fiber, water stain, scratch etc., these beauty defects all can impact the performance of optical system, and dissimilar also different on the impact of system with defect size.

A lot of to optical elements of large caliber surface damage detection method at present, but do not possess the automatic identification function of pit and dust, pit and dust plesiomorphism, be not easily distinguishable, and therefore one of pit and dust damage information that to be tester be concerned about is the most very important to pit and classification identification fast and effectively.

Summary of the invention

The present invention is directed to the deficiency of current detection function, a kind of optical elements of large caliber surface dirt and pit recognition methods are provided, the present invention is by gathering the light field image of optical element surface, to image analysis processing, utilize mode identification method fast dust and pit are effectively classified and identified, accuracy is up to more than 98%.

Optical elements of large caliber surface dirt provided by the invention and pit recognition methods comprise the following steps:

Step 1: utilize image collecting device to gather optical elements of large caliber surface damage image as sample image under light field condition, mark the damage type that each sample image is corresponding simultaneously;

Step 2: extract proper vector for the sample image collecting;

Step 3: the proper vector obtaining is carried out to pre-service;

Step 4: the input variable using pretreated proper vector as pattern recognition classifier device, corresponding damage type label, as the output variable of described pattern recognition classifier device, obtains the mapping relations between input variable and output variable;

Step 5: the mapping relations that obtain based on described step 4, identify for the damage type of test sample book, and recognition result is exported.

Wherein, described image collecting device comprises microlens and area array CCD camera.

Wherein, in described step 1, adopt the mode capturing sample image of axis light illumination.

Wherein, described proper vector includes but not limited to: gradation of image mean value, gradation of image variance, image gradient parameter, contrast, energy, entropy, homogeneity degree.

Wherein, described pre-service at least comprises normalized.

Wherein, the classification of described pattern recognition classifier device comprises dust and pit.

Wherein, obtain the mapping relations between input variable and output variable according to the regularization least square regression model of coring.

Wherein, suppose n input variable x in given d dimension space ₁, x ₂..., x _nand output variable y ₁, y ₂..., y _n∈ R ^c,, there is a linear transformation W ∈ R in c<d ^{d × c}input variable x is mapped as to output variable y, makes y=W ^tx+b, wherein, b ∈ R ^cfor amount of bias, mapping relations W should meet: wherein, J (W) represents the cost function taking W as variable, and λ represents that value is positive regularization parameter, || W|| _frepresent the F norm of W.

Wherein, in described step 5, first by test sample book x _ibring described mapping relations into, obtain the output vector y corresponding with test sample book _i'=W ^tx _i; Then described regularization least square regression model is carried out to coring, obtain the output vector y after coring _ki'; Then by the output variable obtaining, it is mapped in higher dimensional space; Output vector after finally utilizing nearest neighbor classifier to coring in higher dimensional space is classified, and finally obtains the damage type label of test sample book.

Wherein, described coring function is radial basis function K (x, x '):

$K\left(\underset{\mathrm{RBF}}{x,}{x}^{\&prime;}\right)=e^{(-\frac{{\left|\right|x-x^{\&prime;}\left|\right|}^2}{2{\mathrm{\&sigma;}}^2})}.$

The inventive method fast dust to optical element surface and pit is carried out automatic recognition classification.The present invention is dust and the pit classification based on light field imaging picture, the complex process of avoiding darkfield image to gather, and get rid of the impact of non-linear factor on image acquisition under details in a play not acted out on stage, but told through dialogues condition.Meanwhile, the invention allows for a kind of new sorting algorithm: KRR algorithm, compared with other 3 kinds of algorithm for pattern recognitions of commonly using, the present invention has the highest classification accuracy rate.

Brief description of the drawings

Fig. 1 is the process flow diagram of optical elements of large caliber surface dirt of the present invention and pit recognition methods.

Fig. 2 is the pit of collection in the inventive method step 1 and the light field example images of dust, and wherein, Fig. 2 is a) dust light field image, and Fig. 2 is b) pit light field image.

Embodiment

For making the object, technical solutions and advantages of the present invention clearer, below in conjunction with specific embodiment, and with reference to accompanying drawing, the present invention is described in more detail.

Fig. 1 is the process flow diagram of optical elements of large caliber surface dirt of the present invention and pit recognition methods, and as shown in Figure 1, described optical elements of large caliber surface dirt and pit recognition methods comprise the following steps:

Step 1: utilize image collecting device to gather optical elements of large caliber surface damage image as sample image under light field condition, mark the damage type that each sample image is corresponding simultaneously;

In an embodiment of the present invention, described image collecting device comprises microlens and area array CCD camera, and as shown in Figure 2, wherein, Fig. 2 is a) dust light field image to the light field example images of the dust collecting and pit, and Fig. 2 is b) pit light field image.It should be noted that the mode that need adopt axis light illumination in the time of capturing sample image, make visual field bright, instead of utilize scattering of light effect.

Step 2: extract proper vector for the sample image collecting;

In an embodiment of the present invention, described proper vector includes but not limited to following 7 kinds of characteristic of divisions: gradation of image mean value, gradation of image variance, image gradient parameter, contrast, energy, entropy, homogeneity degree, wherein:

The account form of gradation of image mean value is:

Wherein, f (i, j) is the gray-scale value of image mid point (i, j), and n is the number of pixel in image.

The account form of gradation of image variance is:

$\mathrm{Var}=\frac{1}{n}\underset{i}{\mathrm{\&Sigma;}}\underset{j}{\mathrm{\&Sigma;}}[f(i,j)-\mathrm{Mean}{]}^2,$

The account form of image gradient parameter is:

$\mathrm{Gra}=1/\mathrm{MN}\underset{i=i}{\overset{M}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}g(x,y),$

Wherein, the gray-scale value that g (x, y) is the edge gray-scale map mid point (x, y) that obtains through Sobel operator operation, M and N are respectively line number and the columns of image.

The account form of contrast is:

$\mathrm{Con}=\underset{i}{\mathrm{\&Sigma;}}\underset{j}{\mathrm{\&Sigma;}}{(i-j)}^{2}P(i,j),$

Wherein, P is gray level co-occurrence matrixes.

The account form of entropy is:

$\mathrm{Asm}=\underset{i}{\mathrm{\&Sigma;}}\underset{j}{\mathrm{\&Sigma;}}{(i-j)}^{2}P(i,j),$

The account form of energy is:

$\mathrm{Ent}=-\underset{i}{\mathrm{\&Sigma;}}\underset{j}{\mathrm{\&Sigma;}}P(i,j)\log P(i,j),$

The account form of homogeneity degree is:

$\mathrm{Hom}=\underset{i}{\mathrm{\&Sigma;}}\underset{j}{\mathrm{\&Sigma;}}P(i,j)\frac{1}{1+{(i-j)}^2}.$

Wherein, the textural characteristics of considering damage image is random grain feature, do not there is obvious directivity, therefore first obtain the gray level co-occurrence matrixes of 0 °, 45 °, 90 °, 135 ° 4 directions, calculate contrast, energy, entropy, these 4 texture characteristic amounts of homogeneity degree by gray level co-occurrence matrixes again, finally the characteristic quantity calculating in 4 directions is averaged, the mean value of each textural characteristics value is as final textural characteristics value.

Step 3: the proper vector obtaining is carried out to pre-service;

In an embodiment of the present invention, described pre-service at least comprises normalized:

$\stackrel{\&OverBar;}{x_i}=\frac{x_i-\min \left(x_i\right)}{\max \left(x_i\right)-\min \left(x_i\right)},$

Wherein, x _irepresent pending proper vector, represent the proper vector after normalization, x _i∈ [0,1].

Step 4: the input variable using pretreated proper vector as pattern recognition classifier device, corresponding damage type label, as the output variable of described pattern recognition classifier device, obtains the mapping relations between input variable and output variable;

Wherein, the classification of described pattern recognition classifier device comprises dust and pit two classes.

In an embodiment of the present invention, obtain the mapping relations between input variable and output variable according to the regularization least square regression model (KRR) of coring, by proper vector and the damage type label of sample image, the optimal transformation W that calculates regularization least square regression model, obtains the mapping relations between input variable and output variable.

Wherein, the calculating of optimal transformation W is specially: suppose n input variable x in given d dimension space ₁, x ₂..., x _nand output variable y ₁, y ₂..., y _n∈ R ^c,, there is a linear transformation W ∈ R in c<d ^{d × c}input variable x is mapped as to output variable y, makes y=W ^tx+b, wherein, b ∈ R ^cfor amount of bias, optimal transformation W should meet so: wherein, J (W) represents the cost function taking W as variable, and λ represents that value is positive regularization parameter, || W|| _frepresent the F norm of W.

Step 5: the mapping relations that obtain based on described step 4, identify for the damage type of test sample book, and recognition result is exported.

In this step, first by test sample book x _ibring described mapping relations into, obtain the output vector y corresponding with test sample book _i'=W ^tx _i; Then described regularization least square regression model is carried out to coring, obtain the output vector y after coring _ki', in an embodiment of the present invention, the coring function of employing is radial basis function then by the output variable obtaining, it is mapped in higher dimensional space; Finally in higher dimensional space, utilize nearest neighbor classifier (KNN) to classify to the output vector after coring, finally obtain the damage type label of test sample book.

Wherein, the parameter σ in described radial basis function _ibe chosen for standard deviation, choose the σ that classification accuracy rate is the highest _ivalue:

${\mathrm{\&sigma;}}_i=\frac{{\mathrm{\&sigma;}}_0}{2^i}(i=-3,-2,...,\mathrm{2,3}),$

Wherein, σ ₀for standard deviation.

In order to verify validity of the present invention, having chosen 400 groups of data trains, each the light field image pattern gathering is calculated to 7 characteristic of divisions according to step 2, after characteristic of division is normalized, brings in the KRR sorter in step 3 and carry out sorter training, and choose 200 groups of test datas and carry out class test, table one is depicted as the sample number that training and testing sorter gathers.

Table one

Table two is for choosing parameter σ in different coring functions _ithe accuracy of time classification, as can be seen from Table II, the parameter σ in coring function _ishould get

Table two

From the result of above-mentioned experiment, accuracy of the present invention is very high, can be applied to engineering application completely.Table three is the classification accuracy rate contrast of the mode identification method conventional with other three kinds, and as can be seen from Table III, with respect to these three kinds of conventional mode identification methods, the accuracy of the inventive method is the highest.

Table three

Above-described specific embodiment; object of the present invention, technical scheme and beneficial effect are further described; institute is understood that; the foregoing is only specific embodiments of the invention; be not limited to the present invention; within the spirit and principles in the present invention all, any amendment of making, be equal to replacement, improvement etc., within all should being included in protection scope of the present invention.

Claims (10)

1. optical elements of large caliber surface dirt and a pit recognition methods, is characterized in that, the method comprises the following steps:

Step 1: utilize image collecting device to gather optical elements of large caliber surface damage image as sample image under light field condition, mark the damage type that each sample image is corresponding simultaneously;

Step 2: extract proper vector for the sample image collecting;

Step 3: the proper vector obtaining is carried out to pre-service;

Step 4: the input variable using pretreated proper vector as pattern recognition classifier device, corresponding damage type label, as the output variable of described pattern recognition classifier device, obtains the mapping relations between input variable and output variable;

Step 5: the mapping relations that obtain based on described step 4, identify for the damage type of test sample book, and recognition result is exported.

2. method according to claim 1, is characterized in that, described image collecting device comprises microlens and area array CCD camera.

3. method according to claim 1, is characterized in that, in described step 1, adopts the mode capturing sample image of axis light illumination.

4. method according to claim 1, is characterized in that, described proper vector includes but not limited to: gradation of image mean value, gradation of image variance, image gradient parameter, contrast, energy, entropy, homogeneity degree.

5. method according to claim 1, is characterized in that, described pre-service at least comprises normalized.

6. method according to claim 1, is characterized in that, the classification of described pattern recognition classifier device comprises dust and pit.

7. method according to claim 1, is characterized in that, obtains the mapping relations between input variable and output variable according to the regularization least square regression model of coring.

8. method according to claim 7, is characterized in that, supposes n input variable x in given d dimension space ₁, x ₂..., x _nand output variable y ₁, y ₂..., y _n∈ R ^c,, there is a linear transformation W ∈ R in c<d ^{d × c}input variable x is mapped as to output variable y, makes y=W ^tx+b, wherein, b ∈ R ^cfor amount of bias, mapping relations W should meet: wherein, J (W) represents the cost function taking W as variable, and λ represents that value is positive regularization parameter, || W|| _frepresent the F norm of W.

9. method according to claim 7, is characterized in that, in described step 5, first by test sample book x _ibring described mapping relations into, obtain the output vector y corresponding with test sample book _i'=W ^tx _i; Then described regularization least square regression model is carried out to coring, obtain the output vector y after coring _ki'; Then by the output variable obtaining, it is mapped in higher dimensional space; Output vector after finally utilizing nearest neighbor classifier to coring in higher dimensional space is classified, and finally obtains the damage type label of test sample book.

10. method according to claim 9, is characterized in that, described coring function is radial basis function K (x, x '):

$K\left(\underset{\mathrm{RBF}}{x,}{x}^{\&prime;}\right)=e^{(-\frac{{\left|\right|x-x^{\&prime;}\left|\right|}^2}{2{\mathrm{\&sigma;}}^2})}.$