CN105973910A - Structure and texture characteristic-based lamp tube quality detection and flaw classification method and system thereof - Google Patents

Abstract

The invention discloses a structure and texture characteristic-based lamp tube quality detection and flaw classification method and a system thereof. The method comprises the following steps: using a sliding window and a texture characteristic extraction technology to obtain the flaw position of a lamp tube; segmenting flaw images to be classified and various pre-collected flaw sample images to form n subwindows, and calculating the ST descriptor of every subwindow; and respectively calculating the subwindow-class similarity of every subwindow and the class of an image to be detected by using the calculated ST descriptor in order to calculate the image-class similarity, and classifying flaw images to be classified as a maximum image-class similarity class. The finished energy saving lamp tube quality detection method has the advantages of high detection efficiency, high classification precision, and great reduction of the labor intensity of detection personnel.

Description

Lamp tube quality based on structural texture feature detection and defect classification method and system

Technical field

The invention belongs to product quality detection technique field, be specifically related to a kind of lamp tube quality based on structural texture feature Detection and defect classification method and system.

Background technology

During producing power saving fluorescent lamps, need layer of fluorescent powder to be sprayed on equably in electricity-saving lamp tube wall, fluorescent material Coating quality directly influences the using effect of electricity-saving lamp.If making fluorescent coating produce due to technique or accidentalia when producing The flaws such as scuffing, macula lutea, gas line, these substandard products fluorescent tubes will be unable to dispatch from the factory, and needs to carry out different flow process according to different flaw classifications Reprocessing.Therefore the product quality of electricity-saving lamp is detected most important by surface of the light tube Defect Detection and classification.

The product quality detection process of power saving fluorescent lamps is required to enter power saving fluorescent lamps surface blemish situation rapidly and accurately Row detection and classification, on-line checking and the taxonomic methods of real-time high-efficiency detect for quick power saving fluorescent lamps streamline end product quality Extremely important.But, due to factors such as product processing techniques, the quality testing of finished product power saving fluorescent lamps becomes product quality certainly One technical barrier of dynamic detection field.

The Defect Detection of power saving fluorescent lamps and under being sorted in existing process conditions and technical background, general uses the artificial meat of off-line Eye detects and classifies.Under high light, estimated whether fluorescent tube exists flaw by operator's wear dark glasses, and empirically method is true Determine flaw classification, although meet detection and classification demand to a certain extent, but manual detection exists Railway Project: one is people During the most dull, uninteresting detection, it is prone to tired out, easily makes mistakes；Two is severe production environment and the inspection of factory Do not stop the injury watching high light attentively to workman's naked eyes during survey, cause detecting work and be difficult to recruitment；Three is that naked eyes detect automaticity Low, production efficiency is low, it is impossible to meet the actual demand that enterprise quickly, is accurately detected.

Visible, simple manual detection classification is wasted time and energy and accuracy is low, poor reliability, and power saving fluorescent lamps manufacturing enterprise compels It is essential and wants surface of the light tube flaw quality automatically to detect and categorizing system.The Chinese invention of Patent No. ZL 201310015434.4 Patent " detection method of a kind of U-shaped tube cell based on machine vision and detecting system thereof " disclose a kind of by image acquisition, Feature extraction, flaw differentiate calculating, and the tested tube cell dusting of final judgement is up-to-standard or defective, and the method and system can Realize the real-time detection of U-shaped tube cell dusting effect, but be used only for judging U-shaped tube cell flaw with or without, it is impossible to further to various Flaw classification is effectively distinguished.

For fluorescent tube current production technology limitation, the present invention takes into full account batch detection, real-time, classification effectiveness With practical factors such as precision, it is achieved that the classification of energy-conservation fluorescent tube tube cell flaw.

Summary of the invention

For the above-mentioned technological deficiency existing for prior art and deficiency, the invention provides based on structural texture feature Lamp tube quality detection and defect classification method, it is achieved that detection and the district of flaw type in real time of tube cell finished surface flaw Point, detection efficiency and nicety of grading are high, improve the efficiency of power saving fluorescent lamps production line, greatly reduce testing staff simultaneously Labor intensity.

Invention also provides the detection of a kind of lamp tube quality based on structural texture feature and defect classification system, utilize This system, can detect qualified fluorescent tube and flaw fluorescent tube the most efficiently, and can realize the classification to flaw.

The detection of a kind of lamp tube quality based on structural texture feature and defect classification method, including:

(1) flaw of all categories is gathered one or more sample images, set up Sample Storehouse；Specific implementation process is generally: Collect power saving fluorescent lamps in advance and produce the fluorescent tube substandard products that company man's work point is sorted out, substandard products fluorescent tube is carried out people work area according to flaw kind Point, fluorescent tube flaw is classified as several typical categories, flaw of all categories is gathered some images as Sample Storehouse；Make each Classification is C_j, j=1 ..., z, wherein z is total classification number, C={C₁,C₂,...,C_z}；

(2) the fluorescent tube image that Real-time Collection is to be detected, obtains image to be detected；

(3) sliding window and texture characteristic extracting method is utilized to detect whether image to be detected exists flaw, if to be detected There is not flaw in image, then is judged as qualified fluorescent tube；If image to be detected exists flaw, calculate in image to be detected and comprise the flaw The area-of-interest (ROI, Region of Interesting) of defect；

(4) intercept the area-of-interest of flaw image, be considered as image to be classified X, for each image to be classified, by it It is divided into n subwindow, each sample image in Sample Storehouse is divided into n subwindow, each sample image is considered as n The set of individual wicket；N is the natural number more than zero；

(5) it is that all subwindows that step (4) obtains calculate ST description；

(6) utilize ST to describe son to calculate respectively with the subwindow of each apoplexy due to endogenous wind for each subwindow of image to be classified (BC, block-to-class) similarity between subwindow-class；

(7) similarity between the figure-class of the sample image calculating image to be classified and each class；

(8) image to be classified is classified as the class that between figure-class, (IC, image-to-class) similarity is maximum；

Repeat step (2)-(8) and realize the quality testing to multiple power saving fluorescent lampss and defect classification.

As preferably, n is generally the natural number of 4～9.Each image to be classified is divided into m*m subwindow, n=m²。

As preferably, in step (5), the method calculating ST description is:

(5-1) subwindow is divided into S row K row wicket region, there are S*K subband, orderFor the sub-window of x-th The pixel grey scale value set of the kth subband of mouth, k=1 ..., S*K, x=1,2 ..., n；S, K are the natural number more than 1；

(5-2) three below ST calculating each subwindow describes sub:

${\mathrm{\μ}}_k^x=E\left\{\right|W_k^x\left|\right\}---\left(1\right)$

${\mathrm{\σ}}_k^x=E\left\{{\left(\right|W_k^x|-{\mathrm{\μ}}_k^x)}^2\right\}---\left(2\right)$

${\mathrm{\ρ}}_{k1}^x=\frac{E\left\{\left(\right|W_k^x|-{\mathrm{\μ}}_k^x)\left(\right|W_1^x|-{\mathrm{\μ}}_l^x)\right\}}{{\mathrm{\σ}}_k^x{\mathrm{\σ}}_l^x},k\≠l,l=1,...,S*K---\left(3\right)$

In above formula, E{} is for seeking expectation computing,The equal value coefficient of subband for the kth subband of x-th subwindow；For The subband coefficient of standard deviation of the kth subband of x-th subwindow；For kth subband in x-th subwindow and the l subband Between cross-correlation coefficient.

As preferably, in step (6), calculate the method for similarity between subwindow-class as follows:

(6-1) similarity between each subwindow of image to be classified and each subwindow of Different categories of samples image is calculated:

$Q(x,y)=\frac{1}{SK}\underset{k=1}{\overset{SK}{\Σ}}{l}_k^{\frac{1}{3}}(x,y)c_k^{\frac{1}{3}}(x,y)r_k^{\frac{1}{3}}(x,y)---\left(4\right)$

l_k(x y) is the average phase of x-th subwindow and the y-th subwindow in Different categories of samples image in image to be classified Like degree:

$l_k(x,y)=\frac{2{\mathrm{\μ}}_k^x{\mathrm{\μ}}_k^y+K_1}{{\left({\mathrm{\μ}}_k^x\right)}^2+{\left({\mathrm{\μ}}_k^y\right)}^2+K_1}---\left(5\right)$

c_k(x y) is the standard deviation of the y-th subwindow in x-th subwindow and Different categories of samples image in image to be classified Similarity:

$c_k(x,y)=\frac{2{\mathrm{\σ}}_k^x{\mathrm{\σ}}_k^y+K_2}{{\left({\mathrm{\σ}}_k^x\right)}^2+{\left({\mathrm{\σ}}_k^y\right)}^2+K_2}---\left(6\right)$

K₁And K₂For the constant much smaller than 1；

r_k(x y) is the cross-correlation of the y-th subwindow in x-th subwindow and Different categories of samples image in image to be classified Similarity；

$r_k(x,y)=\frac{1}{S+K-2}\[\underset{k\≠l}{\Σ}(1-0.5|{\mathrm{\ρ}}_{kl}^x-{\mathrm{\ρ}}_{kl}^y\left|\right)\]---\left(7\right)$

(6-2) similarity between subwindow-class is drawn:

$Q(x,C)=\underset{y\∈C}{\max }Q(x,y)---\left(8\right)$

The equal value coefficient of subband for the kth subband of Different categories of samples image y-th subwindow；For Different categories of samples figure Subband coefficient of standard deviation as the kth subband of y-th subwindow；For kth in Different categories of samples image y-th subwindow Subband and the cross-correlation coefficient of l intersubband.

As preferably, wherein K₁=0.01～0.02, K₂=0.03～0.05.

As preferably, between the figure-class of the sample image calculating image to be classified and each class, the method for similarity is as follows:

$Q(X,C_j)={\Π}_{i=1}^{n}Q(x_i,C_j)---\left(9\right)$

Q(X,C_j) be image to be classified and each class sample image figure-class between similarity；∏ is that every company takes advantage of fortune Calculate.

Image to be classified classifies as the class that similarity between figure-class is maximum the most at last:

$\hat{C}={\mathrm{argmax}}_{C}Q(X,C).$

One detects and defect classification system based on textural characteristics fluorescent tube end product quality, including:

Image acquisition units, for Real-time Collection power saving fluorescent lamps to be measured image, obtains image to be detected；

Graphics processing unit, for processing described image to be detected, and calculates fluorescent tube defect classification result, Particularly as follows:

I () utilizes sliding window and texture characteristic extracting method to detect whether image to be detected exists flaw, if to be detected There is not flaw in image, then is judged as qualified fluorescent tube；If image to be detected exists flaw, calculate in image to be detected and comprise the flaw The area-of-interest of defect；

(ii) intercept the area-of-interest of flaw image, be considered as image to be classified, for each image to be classified, by it It is divided into n subwindow, each sample image in Sample Storehouse is divided into n subwindow, each sample image is considered as n The set of individual wicket；N is the natural number more than zero；

(iii) it is that all subwindows that step (ii) obtains calculate ST description；

(iv) utilize ST to describe son to count respectively with the subwindow of each apoplexy due to endogenous wind for each subwindow of image to be classified Similarity between operator window-class；

Similarity between the figure-class of v sample image that () calculates image to be classified and each class；

(vi) image to be classified is classified as the class that between figure-class, similarity is maximum.

As further preferably, also include testing result display unit, be used for showing graphics processing unit testing result.

Described graphics processing unit is industrial control computer, realizes power saving fluorescent lamps by sorting algorithm and software programming The classification of surface blemish.Meanwhile, for ease of the monitoring of lamp tube quality detection process, can also include in described graphics processing unit Human interface software, for the power saving fluorescent lamps flaw image to be sorted described in display in real time and defect classification result, record point Class historical data, and the operational order receiving user carries out parameter setting to software.

The present invention passes through high-speed industrial video camera, coordinates the power saving fluorescent lamps image on backlight Real-time Collection streamline, profit The methods such as son are described, it is achieved power saving fluorescent lamps product quality Fast Classification with subwindow ST, have do not contact, not damaged, continuously, real Time, advantage that precision is high；Corresponding testing cost is greatly reduced on the premise of ensureing certainty of measurement.

Accompanying drawing explanation

Fig. 1 is that lamp tube quality based on the structural texture feature detection of the present invention is shown with the steps flow chart of defect classification method It is intended to.

Fig. 2 is several typical case's flaw classification schematic diagrams in embodiment.

Fig. 3 is similarity schematic diagram between Defect Detection figure-class in embodiment.

Detailed description of the invention

In order to more specifically describe the present invention, below in conjunction with the accompanying drawings and detailed description of the invention is to technical scheme It is described in detail.

As it is shown in figure 1, the detection of a kind of lamp tube quality based on structural texture feature and defect classification method, including walking as follows Rapid:

(1) collect power saving fluorescent lamps in advance and produce the power saving fluorescent lamps substandard products that company man's work point is sorted out, to substandard products fluorescent tube according to the flaw Defect kind is manually distinguished, and fluorescent tube flaw classifies as several typical categories (such as Fig. 2), and making each classification is C_j, j= 1 ..., z, C={C₁,C₂,...,C_zWherein z be total classification number, for the natural number more than zero；Flaw of all categories is gathered one Determined number sample image (the big I of picture specification determines based on experience value) is as Sample Storehouse；

(2) the power saving fluorescent lamps image that Real-time Collection is to be detected；

(3) sliding window and texture characteristic extracting method is utilized to detect whether above-mentioned fluorescent tube image exists flaw, if to be checked There is flaw in altimetric image, calculates area-of-interest (ROI, the Region of comprising flaw in fluorescent tube image Interesting)；

(4) intercept the area-of-interest part of flaw image, be considered as image to be classified X, be divided into n subwindow, By the same way, each sample image in Sample Storehouse is divided into n subwindow, each sample image is considered as n The set of wicket.N=m², m is the positive integer more than 1；M is generally 2 or 3；

(5) subwindow is divided into a S row K row wicket region, and (value of S, K, to there are S*K subband After the size about 10*10 pixel size of subband that obtains be as the criterion), orderPixel for the kth subband of x-th subwindow Gray value set, k=1 ..., S*K, x=1,2 ..., m²；In formula (1), E{} is for seeking expectation computing.Obtain for step (4) All subwindows calculate three below ST and describe son；

I. the equal value coefficient of subband

${\mathrm{\μ}}_k^x=E\left\{\right|W_k^x\left|\right\}---\left(1\right)$

II. subband coefficient of standard deviation

${\mathrm{\σ}}_k^x=E\left\{{\left(\right|W_k^x|-{\mathrm{\μ}}_k^x)}^2\right\}---\left(2\right)$

III. kth subband and the cross-correlation coefficient of l intersubband

${\mathrm{\ρ}}_{kl}^x=\frac{E\left\{\left(W_k^x\right|-{\mathrm{\μ}}_k^x)\left(W_1^x\right|-{\mathrm{\μ}}_l^x)\right\}}{{\mathrm{\σ}}_k^x{\mathrm{\σ}}_l^x},k\≠l,l=1,\mathrm{...},S*K---\left(3\right)$

The equal value coefficient of subband for the kth subband of x-th subwindow；Kth subband for x-th subwindow Subband coefficient of standard deviation；For kth subband and the cross-correlation coefficient of l intersubband in x-th subwindow；

(6) utilize above-mentioned ST to describe each subwindow that son is image to be classified and calculate sub-window respectively with each class (BC, block-to-class) similarity between mouth-class.First each subwindow and the Different categories of samples figure of image to be classified are calculated Similarity between each subwindow of picture:

$Q(x,y)=\frac{1}{SK}\underset{k=1}{\overset{SK}{\Σ}}{l}_k^{\frac{1}{3}}(x,y)c_k^{\frac{1}{3}}(x,y)r_k^{\frac{1}{3}}(x,y)---\left(4\right)$

Wherein:

l_k(x y) is the average phase of x-th subwindow and the y-th subwindow in Different categories of samples image in image to be classified Like degree

$l_k(x,y)=\frac{2{\mathrm{\μ}}_k^x{\mathrm{\μ}}_k^y+K_1}{{\left({\mathrm{\μ}}_k^x\right)}^2+{\left({\mathrm{\μ}}_k^y\right)}^2+K_1}---\left(5\right)$

c_k(x y) is the standard deviation of the y-th subwindow in x-th subwindow and Different categories of samples image in image to be classified Similarity

$c_k(x,y)=\frac{2{\mathrm{\σ}}_k^x{\mathrm{\σ}}_k^y+K_2}{{\left({\mathrm{\σ}}_k^x\right)}^2+{\left({\mathrm{\σ}}_k^y\right)}^2+K_2}---\left(6\right)$

K in formula (5) and (6)₁And K₂For the constant much smaller than 1, typically take K₁=0.01, K₂=0.03；

r_k(x y) is the cross-correlation of the y-th subwindow in x-th subwindow and Different categories of samples image in image to be classified Similarity

$r_k(x,y)=\frac{1}{S+K-2}\[\underset{k\≠l}{\Σ}(1-0.5|{\mathrm{\ρ}}_{kl}^x-{\mathrm{\ρ}}_{kl}^y\left|\right)\]---\left(7\right)$

Finally draw similarity between subwindow-class:

$Q(x,C)=\underset{y\∈C}{\max }Q(x,y)---\left(8\right)$

Y is a subwindow of an apoplexy due to endogenous wind, y=1,2 ..., n；Son for the kth subband of y-th subwindow Carry equal value coefficient；Subband coefficient of standard deviation for the kth subband of y-th subwindow；For kth in y-th subwindow Subband and the cross-correlation coefficient of l intersubband.Above-mentioned various in " | | " be and seek amplitude operator.

(7) (IC, image-to-class) similarity (such as Fig. 3) between the figure-class of calculating image to be classified and each class:

$Q(X,C_j)={\Π}_{i=1}^{n}Q(x_i,C_j)---\left(9\right)$

Wherein ∏ is every even multiplication；

(8) image to be classified is classified as the class that similarity between figure-class is maximum:

$\hat{C}={\mathrm{argmax}}_{C}Q(X,C)---\left(10\right)$

Wherein arg is for asking subscript computing；Maximum Q (X, C is calculated by formula (10)_j) time the subscript of C, finally give The classification number of the class that similarity is maximum between figure-class, and then confirm final classification；Repeat step (3)～(8) to realize multiple The quality testing of power saving fluorescent lamps and defect classification.

Fig. 2 and Fig. 3 is should a specific embodiment in aforementioned manners:

Fig. 2 is the sample image of several typical categories, the i-th class-scuffing, the i-th i class-scraping, the i-th ii class-gas line, the i-th v Class-duck eye, v class-inequality, vi class-shedding；

Fig. 3 middle graph is image to be classified, through being finally calculated between the figure-class of image to be classified and each class (IC, image-to-class) similarity is: and between the figure-class of the i-th class, similarity is 0.7268, and between the figure-class of the i-th i class Similarity is 0.8048, and between the figure-class of the i-th ii class, similarity is 0.6025, and between the figure-class of the i-th v class, similarity is Between 0.7395, and the figure-class of v class, similarity is that between 0.7231, and the figure-class of vi class, similarity is 0.7863.Through commenting Valency, playback the i-th i class by the image to be classified of middle graph, scratches.By further eye-observation, the inspection of present embodiment The surface of the light tube defect classification result that survey method draws has higher efficiency and reliability, credible.

It is also possible to exist the to be sorted of multiple different classes of or identical category band flaw image on certain fluorescent tube Image, for there is the fluorescent tube of multiple flaw, can individually classify to each next time, it is determined that method etc. are identical with above-mentioned principle.

This fluorescent tube defect classification system includes image acquisition units, graphics processing unit and classification results display unit.

Image acquisition units, for Real-time Collection power saving fluorescent lamps to be measured image；Industrial camera, camera lens and photograph can be used Source, Mingguang City realizes.Industrial camera uses Daheng's 1/3 inch of cmos camera of DH-GV400UM black and white, full frame exposure scanning side Formula, resolution is 752 × 480, and frame per second is up to 60 frames/second, and output interface is USB, and camera lens bayonet socket is C/CS mouth, compact, It is easily installed, real-time testing requirement can be met；The undistorted camera lens of 8mm of Japan Computar selected by camera lens.Light source selects latitude bright 160 × 160mm face shape blue led light source, light source controller selects latitude bright VL-LC-11-4CH model USB light source controller.

Graphics processing unit, for processing described power saving fluorescent lamps image, and calculates fluorescent tube defect classification knot Really；Graphics processing unit hardware can use industrial control computer, and core is the computer software write.Specific works side Formula is:

I (), in the power saving fluorescent lamps image that image acquisition units gathers, utilizes sliding window and texture characteristic extracting method Obtain fluorescent tube flaw position；

(ii) all kinds of flaw sample images of flaw image to be sorted and pre-collecting are divided into n subwindow, and for institute There is subwindow to calculate ST and describe son；

(iii) utilize the ST calculated to describe each subwindow that son is image to be detected and calculate son respectively with all classes Similarity between window-class, and then calculate similarity between figure-class, image to be classified classifies as similarity maximum between figure-class the most at last A class.

Described graphics processing unit realizes power saving fluorescent lamps surface blemish by software programming and above-mentioned sorting algorithm Classification.Meanwhile, for ease of the monitoring of lamp tube quality detection process, in described graphics processing unit, can also include that man machine interface is soft Part, for the power saving fluorescent lamps flaw image to be sorted described in real time display and defect classification result, record sort historical data, And the operational order receiving user carries out parameter setting to software.Real process is shown by display screen.Industry Control calculates Machine connects industrial camera and light source controller by USB data line.Industrial control computer uses and grinds China's Industry Control calculating Machine, this machine uses Intel dual core processor, dominant frequency 3.0GHz, internal memory 2G, hard disk 160G, 19 cun of liquid crystal displays, meets industry The requirement of on-the-spot adverse circumstances.

Testing result display unit, the actually optional display being connected with industrial control computer, cooperation is write Product defect categorizing system monitoring software, can show graphics processing unit classification results intuitively.

Claims (8)

1. lamp tube quality based on a structural texture feature detection and defect classification method, it is characterised in that including:

(1) flaw of all categories is gathered one or more sample images, set up Sample Storehouse；Making each classification is C_j, j=1 ..., Z, wherein z is total classification number, C={C₁,C₂,…,C_z}；

(2) the fluorescent tube image that Real-time Collection is to be detected, obtains image to be detected；

(3) sliding window and texture characteristic extracting method is utilized to detect whether image to be detected exists flaw, if image to be detected There is not flaw, be then judged as qualified fluorescent tube；If image to be detected exists flaw, calculate in image to be detected and comprise flaw Area-of-interest；

(4) intercept the area-of-interest of flaw image, be considered as image to be classified X, for each image to be classified, split For n subwindow, each sample image in Sample Storehouse is divided into n subwindow, each sample image is considered as n individual little The set of window；N is the natural number more than zero；

(5) it is that all subwindows that step (4) obtains calculate ST description；

(6) utilize ST to describe son and calculate sub-window respectively for each subwindow of image to be classified and the subwindow of each apoplexy due to endogenous wind Similarity between mouth-class；

(7) similarity between the figure-class of the sample image calculating image to be classified and each class；

(8) image to be classified is classified as the class that between figure-class, similarity is maximum；

Repeat step (2)-(8) and realize the quality testing to multiple power saving fluorescent lampss and defect classification.

Lamp tube quality based on structural texture feature the most according to claim 1 detection and defect classification method, its feature Being, n is the natural number of 4～9.

Lamp tube quality based on structural texture feature the most according to claim 1 detection and defect classification method, its feature Being, the method calculating ST description is:

(5-1) subwindow is divided into S row K row wicket region, there are S*K subband, orderFor x-th subwindow The pixel grey scale value set of kth subband, k=1 ..., S*K, x=1,2 ..., n；S, K are the natural number more than 1；

(5-2) three below ST calculating each subwindow describes sub:

${\mathrm{\μ}}_k^x=E\left\{\right|W_k^x\left|\right\}---\left(1\right)$

${\mathrm{\σ}}_k^x=E\left\{{\left(\right|W_k^x|-{\mathrm{\μ}}_k^x)}^2\right\}---\left(2\right)$

${\mathrm{\ρ}}_{kl}^x=\frac{E\left\{\left(\right|W_k^x|-{\mathrm{\μ}}_k^x)\left(\right|W_l^x|-{\mathrm{\μ}}_l^x)\right\}}{{\mathrm{\σ}}_k^x{\mathrm{\σ}}_l^x},k\≠l,l=1,...,S*K---\left(3\right)$

In above formula, E{} is for seeking expectation computing；The equal value coefficient of subband for the kth subband of x-th subwindow；For x-th The subband coefficient of standard deviation of the kth subband of subwindow；Mutual for kth subband in x-th subwindow and l intersubband Correlation coefficient.

Lamp tube quality based on structural texture feature the most according to claim 3 detection and defect classification method, its feature It is, calculates the method for similarity between subwindow-class as follows:

(6-1) similarity between each subwindow of image to be classified and each subwindow of Different categories of samples image is calculated:

$Q(x,y)=\frac{1}{SK}\underset{k=1}{\overset{SK}{\Σ}}{l}_k^{\frac{1}{3}}(x,y)c_k^{\frac{1}{3}}(x,y)r_k^{\frac{1}{3}}(x,y)---\left(4\right)$

l_k(x, y) is the average similarity of the y-th subwindow in x-th subwindow and Different categories of samples image in image to be classified:

$l_k(x,y)=\frac{2{\mathrm{\μ}}_k^x{\mathrm{\μ}}_k^y+K_1}{{\left({\mathrm{\μ}}_k^x\right)}^2+{\left({\mathrm{\μ}}_k^y\right)}^2+K_1}---\left(5\right)$

c_k(x is y) that in image to be classified, x-th subwindow is similar to the standard deviation of the y-th subwindow in Different categories of samples image Degree:

$c_k(x,y)=\frac{2{\mathrm{\σ}}_k^x{\mathrm{\σ}}_k^y+K_2}{{\left({\mathrm{\σ}}_k^x\right)}^2+{\left({\mathrm{\σ}}_k^y\right)}^2+K_2}---\left(6\right)$

K₁And K₂For the constant much smaller than 1；

r_k(x is y) that in image to be classified, x-th subwindow is similar to the cross-correlation of the y-th subwindow in Different categories of samples image Degree；

$r_k(x,y)=\frac{1}{S+K-2}\[\underset{k\≠l}{\Σ}(1-0.5|{\mathrm{\ρ}}_{kl}^x-{\mathrm{\ρ}}_{kl}^y\left|\right)\]---\left(7\right)$

(6-2) similarity between subwindow-class is drawn:

$Q(x,C)=\underset{y\∈C}{max}Q(x,y)---\left(8\right)$

The equal value coefficient of subband for the kth subband of Different categories of samples image y-th subwindow；For Different categories of samples image y The subband coefficient of standard deviation of the kth subband of individual subwindow；For kth subband in Different categories of samples image y-th subwindow and The cross-correlation coefficient of l intersubband.

Lamp tube quality based on structural texture feature the most according to claim 4 detection and defect classification method, its feature It is, K₁=0.01～0.02, K₂=0.03～0.05.

Lamp tube quality based on structural texture feature the most according to claim 4 detection and defect classification method, its feature Being, between the figure-class of the sample image calculating image to be classified and each class, the method for similarity is as follows:

$Q(X,C_j)={\mathrm{\Π}}_{i=1}^{n}Q(x_i,C_j)---\left(9\right)$

Q(X,C_j) be image to be classified and each class sample image figure-class between similarity；∏ is every even multiplication.

7. lamp tube quality based on a structural texture feature detection and defect classification system, it is characterised in that including:

Image acquisition units, for Real-time Collection fluorescent tube to be measured image, obtains image to be detected；

Graphics processing unit, for processing described image to be detected, and calculates fluorescent tube defect classification result, specifically For:

I () utilizes sliding window and texture characteristic extracting method to detect whether image to be detected exists flaw, if image to be detected There is not flaw, be then judged as qualified fluorescent tube；If image to be detected exists flaw, calculate in image to be detected and comprise flaw Area-of-interest；

(ii) intercept the area-of-interest of flaw image, be considered as image to be classified, for each image to be classified, split For n subwindow, each sample image in Sample Storehouse is divided into n subwindow, each sample image is considered as n individual little The set of window；N is the natural number more than zero；

(iii) it is that all subwindows that step (ii) obtains calculate ST description；

(iv) utilize ST to describe son and calculate son respectively for each subwindow of image to be classified and the subwindow of each apoplexy due to endogenous wind Similarity between window-class；

Similarity between the figure-class of v sample image that () calculates image to be classified and each class；

(vi) image to be classified is classified as the class that between figure-class, similarity is maximum.

Lamp tube quality based on structural texture feature the most according to claim 7 detection and defect classification system, its feature It is, also includes testing result display unit.