CN101719273A - On-line self-adaptation extraction method of metallurgy strip surface defect based on one-dimension information entropy - Google Patents

Abstract

The invention relates to an on-line self-adaptation extraction method of metallurgy strip surface defect based on one-dimension information entropy. The on-line self-adaptation extraction method comprises the following steps of: firstly, carrying out Gaussian smoothing on an acquired source image; then carrying out spatial domain enhancement on the smoothed image in specific to defect objects; then carrying out defect object segmentation on the processed image by utilizing a self-adaption entropy method; and then, carrying out a closing operation on a segmented binary image by using a method in morphology for improving the connectivity of the binary image to obtain a final defect image. The invention has the advantages of precision defect location, high efficiency, small occupied resources, and simple and stable as well as reliable operation, provides guarantee for the precision of consequent detection technology, effectively solves the problems of extraction invalidation, easy breaking-down of a system data bus and the like when a traditional strip surface quality detection system based on a gray-scale abnormity method encounters large amount of benign defects and strip pattern interferences, and greatly enhances effective object extraction rate.

Description

Based on one-dimension information entropy metallurgy strip surface defect online adaptive extracting method

Technical field

The present invention relates to a kind ofly, be used for the strip surface quality and detect based on one-dimension information entropy metallurgy strip surface defect online adaptive extracting method.

Background technology

In the metallurgical processing industry, along with improving constantly of production capacity, the demand that detects for product surface quality constantly increases both at home and abroad, and effect of requirement is more and more higher for detecting.No matter at present domestic be for the equipment of introducing or for the experimental checkout equipment of developing voluntarily, be not fine for the locating effect that detects target all, difference, the template situation that is subjected to material type easily is bad, the influence of environmental interference and a large amount of optimum attachments.

The surface defects detection system is distributed in basic automatization and the Process Control System level two as the key components of quality control system, plays to the management of On-line Product quality testing, finished product and to the effects such as processing process control of material.And the realization of these functions all will with material surface unusual effectively orientate important prerequisite fast as, and this link is limited by the influence of above-mentioned every factor, and therefore how seeking effective defective extracting method is the important channel that surface detecting system improves defects detection and discrimination.

Summary of the invention

The objective of the invention is to overcome the deficiency that prior art exists, provide a kind of based on one-dimension information entropy metallurgy strip surface defect online adaptive extracting method, the defective that effectively solves in the strip surface quality detection system is extracted problem.

Purpose of the present invention is achieved through the following technical solutions:

Based on one-dimension information entropy metallurgy strip surface defect online adaptive extracting method, characteristics are: at first, to carry out Gauss level and smooth to gathering the source images that comes; Then, the image after level and smooth is carried out the enhancing of spatial domain at the defective target; Carrying out the defective target with the method for the adaptive entropy image after to previous processed again cuts apart; Then, the method in the morphology is carried out closed operation to the bianry image method after cutting apart, and improves its connectedness and obtains final defect image; Specifically may further comprise the steps:

1. with smoothing operator the material surface image is carried out convolution algorithm, remove the noise of image;

2. the image after level and smooth is carried out the spatial domain and strengthen, to strengthen target;

3. utilize the method that maximizes one-dimension information entropy that image is carried out defective again and cut apart, obtain the image that a threshold is cut apart;

4. utilize the image after closed operation is cut apart threshold to handle then, remove the isolated point and the noise of defective target internal, improve the connectedness of image-region, obtain final strip surface defect image.

Further, above-mentioned based on one-dimension information entropy metallurgy strip surface defect online adaptive extracting method, wherein, smoothing operator is Gauss's smoothing operator; The territory strengthens adopts Laplace operator; The information entropy that the maximization one-dimension information entropy adopts is an image one dimension histogram information entropy; The closed operation operator is that an all elements is 1 masterplate.

Substantive distinguishing features and obvious improvement that technical solution of the present invention is outstanding are mainly reflected in:

Image noise is removed in the at first level and smooth and spatial domain enhancing by Gauss of method of the present invention, and the target that brightening will be extracted is for image segmentation provides a high-quality image; Then adopt and based on the method for one-dimension information entropy image is carried out defective and cut apart, in the light and shade field picture all is fit to.But because threshold is cut apart is the process of a subjectivity, the bianry image of gained may comprise not to be needed or undesired information, therefore the closed operation operator in the last employing morphology carries out regional connectivity to the image after cutting apart, remove the isolated or noise spot in the target, the target of having broken is linked up, orient final defect map.Defect location is accurate, efficient is high and it is few to take resource, for the precision of subsequent detection technology provides assurance.Method simple and stable of the present invention is reliable, efficiently solve tradition and extract inefficacy, easily cause system data bus problems such as paralysis to occur based on strip surface quality detection system appearance when running into the interference of a large amount of optimum defectives and template of the unusual method of gray scale, the effective target extraction ratio obtains bigger enhancing.

Description of drawings

Fig. 1: the gaussian kernel of the Gaussian filter that the present invention is used;

Fig. 2 a: example that is of a size of 3 gaussian kernel;

Fig. 3 a: example that is of a size of 5 gaussian kernel.

Fig. 4: spatial domain of the present invention strengthens used laplace kernel;

Fig. 5 a: example of laplace kernel;

Fig. 6: image is about gray-scale value one dimension histogram;

Fig. 7: the employed masterplate of closed operation among the present invention.

Fig. 8: process flow diagram of the present invention.

Embodiment

The invention provides a kind of based on one-dimension information entropy metallurgy strip surface defect online adaptive extracting method, at first adopt Gauss's smoothing operator that image is carried out smoothing processing, carry out the image spatial domain at target again and strengthen the material surface view data of coming by the line array sensor collection; Adopt then and based on the method for histogrammic one dimension adaptive entropy image is carried out the defective target and cut apart; Remove the isolated point and the noise of target internal at last with the closed operation in the morphology, improve the connectivity in zone, obtain final defect image; Its flow process is:

1. with smoothing operator the material surface image is carried out convolution algorithm, remove the noise of image; Wherein, smoothing operator is Gauss's smoothing operator;

2. the image after level and smooth is carried out the spatial domain and strengthen, to strengthen target, the territory strengthens adopts Laplace operator; Carry out level and smooth in succession and the spatial domain strengthens to image, smoothing operator is removed the noise of image, and spatial domain enhancement algorithms enhancing target is for the image segmentation of back provides an image that noise is few, with clearly defined objective;

3. utilize the method for maximization one-dimension information entropy that image is carried out defective again and cut apart, obtain the image that a threshold is cut apart, determine the threshold value of image segmentation to obtain split image by the method for maximization image adaptive entropy; Wherein, the information entropy of maximization one-dimension information entropy employing is an image one dimension histogram information entropy;

4. the image after closed operation is cut apart threshold in the imagery exploitation morphology after cutting apart is handled, to obtain the accurate image of defect location, remove the isolated point and the noise of defective target internal, improve the connectedness of image-region, obtain final strip surface defect image; Wherein, adopting closed operation (4) operator is Fig. 6, is that an all elements is 1 masterplate.

Specifically comprise following steps:

1) source images is carried out smoothing processing: the brightness that Gauss smoothly weakens in the neighborhood of pixels changes, and the shape of level and smooth target weakens details, is similar to smoothing operator, but fuzzy influence than smoothing operator a little less than.At this, mainly be to be used to remove some noises, so adopt the more weak Gaussian filter of fuzzy influence;

The level and smooth convolution filter of Gauss is an average filter, the masterplate below nuclear uses, as shown in Figure 1, and a wherein, b, c, d are positive integers, X＞1.The coefficient of the Gaussian convolution nuclear of intended size is preferably as much as possible near the round values on the Gaussian curve.Fig. 2 is an example of gaussian kernel.

2) image after level and smooth being carried out the spatial domain strengthens: adopt laplace kernel as the convolution masterplate, image is carried out convolution algorithm, strengthen the image spatial information (si), reach the purpose that strengthens the defective target, for follow-up defective extraction work provides a high-quality picture, to improve the defect location precision.Laplace kernel is a second order local derviation, and the masterplate that nuclear uses is shown in Figure 4, a wherein, and b, c, d are negative integer, center coefficient X＞2 (| a|+|b|+|c|+|d|).Fig. 5 is an example of gaussian kernel.

3) with the adaptive entropy method image that strengthens after the target is carried out Threshold Segmentation to obtain preliminary defect map:

Fig. 6 is the one dimension histogram of image about gray-scale value, and transverse axis is represented gray-scale value i, and the tonal range of image is [1, L], and the longitudinal axis represents that number h appears in certain gray-scale value i _iSuppose that d is the segmentation threshold of target and background, A district and B district represent target area and background area respectively among the figure, then the one-dimension information entropy of target and background is respectively H (A) and H (B), thereby the histogram information entropy of image is H (d)=H (A)+H (B), and best threshold value is

Utilize the binaryzation function at last

$f\left(i\right)=\left\{\begin{array}{cc}1& i\≤d\\ 0& i>d\end{array}\right.$

The defective target is tentatively split from image, and wherein 1 represents target, 0 expression background, thus the defective target is split from image.

4) by closed operation the bianry image that Threshold Segmentation obtains is further handled: Threshold Segmentation is a target leaching process, and the bianry image that obtains may comprise undesired information, as contact portion of noise, target and image border or the like.Adopt the closed operation in the morphology can remove these unwanted information, influence target shape, thereby improve the information of bianry image.

Technical scheme for a better understanding of the present invention, describes in further detail by example below in conjunction with accompanying drawing.

Fig. 1 is that the present invention adopts the convolution kernel of Gauss's smoothing filter for subsequent treatment provides a less image of noise, and wherein a, b, c, d are integers, X＞1.Fig. 2 and Fig. 3 are examples of gaussian kernel.

Because all coefficients in the gaussian kernel are positive, each pixel is the weighted mean of consecutive point.The weights of consecutive point are big more, and the influence that the new value of central point is produced is big more.

Different with general smoothing kernel is that the center coefficient of gaussian filtering is greater than 1.So the weighted value that the original value of pixel is taken advantage of is greater than the weights of any pixel in the neighborhood.Therefore, the big more corresponding more delicate smoothing effect of center coefficient.The nuclear size is big more, and smooth effect is big more.To be lower than other smoothing operator to the fog-level of image, so when removing noise, farthest kept the information of image.

Fig. 4 is a laplace kernel, is to the present invention is directed to target to extract and a spatial domain enhancing masterplate of design, has manifested the defective target.Fig. 5 is an example of laplace kernel.

Gauss smoothly combines with the spatial domain wild phase, and for next step image segmentation provides a noise few, the tangible image of defective target will improve next step defective segmentation precision virtually.

Fig. 6 is the one dimension histogram of image about gray-scale value, and transverse axis is represented gray-scale value, and the longitudinal axis represents that number h appears in certain gray-scale value i _iIf the tonal range of image is [1, L], in image segmentation, represents to cut apart the back image with information entropy and comprise that quantity of information gets big or small in the original image.The probability that pixel gray-scale value i occurs is

Wherein M * N is the size of image.If d is the segmentation threshold of target and background, establish that A represents that target area and B represent the background area among the figure, then the probability that occurs of target and background is respectively

With

Then the one-dimension information entropy of target and background is respectively:

$H\left(A\right)=-\underset{i=1}{\overset{d}{\mathrm{\Σ}}}{p}_i/P_A{\log }_2(p_i/P_A),$

$H\left(B\right)=-\underset{i=d+1}{\overset{L}{\mathrm{\Σ}}}{p}_i/P_B{\log }_2(p_i/P_B).$

Under threshold value d, the information entropy of image is: H (d)=H (A)+H (B).Will

With

Among substitution H (d)=H (A)+H (B),

$H\left(d\right)=-\frac{1}{\underset{i=1}{\overset{d}{\mathrm{\&Sigma;}}}{h}_i}\underset{i=1}{\overset{d}{\mathrm{\&Sigma;}}}{h}_i{\log }_2{h}_i+{\mathrm{<figure-callout\; id="2"\; label="log"\; filenames="H2009102132343E0000012.png,H2009102132343E0000013.png"\; state="\{\{state\}\}">log</figure-callout>}}_2\underset{i=1}{\overset{d}{\mathrm{\&Sigma;}}}{h}_i+-\frac{1}{\underset{i=d+1}{\overset{L}{\mathrm{\&Sigma;}}}{h}_i}\underset{i=d+1}{\overset{L}{\mathrm{\&Sigma;}}}{h}_i{\log }_2{h}_i+{\log }_2\underset{i=d+1}{\overset{L}{\mathrm{\&Sigma;}}}{h}_i$

Order $a=\underset{i=1}{\overset{d}{\mathrm{\&Sigma;}}}{h}_i,$ Then $\underset{i=d+1}{\overset{L}{\mathrm{\&Sigma;}}}{h}_i=\mathrm{MN}-a,$ So

$H\left(d\right)=-\frac{1}{a}\underset{i=1}{\overset{d}{\mathrm{\&Sigma;}}}{h}_i{\log }_2{h}_i-\frac{1}{\mathrm{MN}-a}\underset{i=d+1}{\overset{L}{\mathrm{\&Sigma;}}}{h}_i{\log }_2{h}_i+{\log }_2\left(a(\mathrm{MN}-a)\right)$

When H (d) obtains the pairing threshold value d of maximum value is the segmentation threshold of being asked, and s and t are respectively the cut-point of S axle and T axle.Image carries out binary conversion treatment according to the segmentation threshold vector, and the binaryzation function is:

$f\left(i\right)=\left\{\begin{array}{cc}1& i\&le;d\\ 0& i>d\end{array}\right.$

Wherein 1 represents target, 0 expression background, thus the defective target is split from image.

The masterplate that Fig. 7 adopts for closed operation among the present invention couples together the defective that disconnects, and removes the heterogeneous point in the defective target simultaneously.Closed operation in the morphology is earlier with masterplate image to be carried out dilation operation in the morphology in fact, carries out erosion operation in the morphology with same masterplate then.

Fig. 8 has illustrated the flow process of whole invention, and its principle is: utilize the level and smooth and spatial domain of Gauss to strengthen source images is carried out pre-service, remove noise, enhancing defective target is for follow-up image segmentation provides a high-quality image.Then determine segmentation threshold with one-dimension information entropy, obtain a defective split image, further remove the heterogeneous point that some are divided into target by mistake again, obtain an image that segmentation precision is high with the closed operation in the morphology by maximization.Each step here interlocks step by step, all is to make every effort to remove as much as possible heterogeneous point, accurately orients defective.

In sum, image noise is removed in the at first level and smooth and spatial domain enhancing by Gauss of method of the present invention, and the target that brightening will be extracted is for image segmentation provides a high-quality image; Then adopt and based on the method for one-dimension information entropy image is carried out defective and cut apart, in the light and shade field picture all is fit to.But because threshold is cut apart is the process of a subjectivity, the bianry image of gained may comprise not to be needed or undesired information, therefore the closed operation operator in the last employing morphology carries out regional connectivity to the image after cutting apart, remove the isolated or noise spot in the target, the target of having broken is linked up, orient final defect map.Defect location is accurate, efficient is high and it is few to take resource, for the precision of subsequent detection technology provides assurance.Method simple and stable of the present invention is reliable, efficiently solve tradition and extract inefficacy, easily cause system data bus problems such as paralysis to occur based on strip surface quality detection system appearance when running into the interference of a large amount of optimum defectives and template of the unusual method of gray scale, the effective target extraction ratio obtains bigger enhancing.

Below only be concrete exemplary applications of the present invention, protection scope of the present invention is not constituted any limitation.All employing equivalents or equivalence are replaced and the technical scheme of formation, all drop within the rights protection scope of the present invention.

Claims (5)

1. based on one-dimension information entropy metallurgy strip surface defect online adaptive extracting method, it is characterized in that: may further comprise the steps:

1. with smoothing operator the material surface image of gathering is carried out convolution algorithm, remove the noise of image;

2. the image after level and smooth is carried out the spatial domain and strengthen, to strengthen target;

3. utilize the method that maximizes one-dimension information entropy that image is carried out defective again and cut apart, obtain the image that a threshold is cut apart;

4. closed operation in the imagery exploitation morphology after cutting apart is handled it, removed the isolated point and the noise of defective target internal, improve the connectedness of image-region, obtain final strip surface defect image.

2. according to claim 1 based on one-dimension information entropy metallurgy strip surface defect online adaptive extracting method, it is characterized in that: smoothing operator is Gauss's smoothing operator.

3. according to claim 1 based on one-dimension information entropy metallurgy strip surface defect online adaptive extracting method, it is characterized in that: the spatial domain strengthens adopts Laplace operator.

4. according to claim 1 based on one-dimension information entropy metallurgy strip surface defect online adaptive extracting method, it is characterized in that: the information entropy that the maximization one-dimension information entropy adopts is an image one dimension histogram information entropy.

5. according to claim 1 based on one-dimension information entropy metallurgy strip surface defect online adaptive extracting method, it is characterized in that: the closed operation operator is that an all elements is 1 masterplate.

Images