CN106093066B - A kind of magnetic tile surface defect detection method based on improved machine vision attention mechanism - Google Patents

Abstract

The invention discloses a magnetic tile surface defect detection method based on an improved machine vision attention mechanism. 2. Divide the obtained image evenly into a*b image blocks, and then use the gray feature quantity of the divided image blocks to distinguish defective image blocks from non-defective image blocks; 3. Use the improved Itti visual attention mechanism model to calculate The saliency of the obtained defect image block is selected from the primary features to form a comprehensive saliency map; 4. The Otsu threshold segmentation algorithm is used to threshold the comprehensive saliency map to extract the defect area. The invention effectively overcomes the problems of uneven brightness, small defect area of magnetic tiles, and texture interference of magnetic tiles by using morphological processing, image segmentation and visual attention mechanism, and can quickly and effectively extract various types of magnetic tiles Defects are highly adaptable.

Description

A Method for Surface Defect Detection of Magnetic Tile Based on Improved Machine Vision Attention Mechanism

technical field

The invention relates to the technical field of machine vision, in particular to a method for detecting surface defects of magnetic tiles based on an improved machine vision attention mechanism.

Background technique

Ferrite magnet tile is a kind of tile-shaped magnet mainly used in permanent magnet motors, and its quality directly affects the overall performance of permanent magnet motors. During the production process of magnetic tiles, due to technological problems, the surface of magnetic tiles is prone to defects such as cracks, damage, pitting, etc., which directly affects the normal use of magnetic tiles. At present, the judgment of the surface defects of magnetic tiles in industrial production basically adopts manual detection, which has poor detection accuracy, low detection efficiency and high labor cost.

With the continuous development of machine vision, defect detection technology based on machine vision has begun to be widely used in the monitoring of the surface quality of industrial products. The use of machine vision automatic detection can improve the production efficiency of enterprises, reduce labor costs, and increase the competitiveness of enterprises.

For magnetic tile products, due to their own characteristics such as inconspicuous grayscale difference and surface curvature, it is easy to cause uneven illumination and low image grayscale contrast, which brings certain difficulties to the development of high-precision and fast-speed magnetic tile detection methods. . In the "Journal of Computer-Aided Design and Graphics", Li Xueqin proposed an automatic detection method for magnetic tile defects using an adaptive threshold surface in the non-subsampled Contourlet domain. This method can ensure a high accuracy of magnetic tile surface defect detection, but Calculation time is long. Yu Yongwei et al. proposed an adaptive linear morphology algorithm to segment defects based on the distribution of gray values on the surface of magnetic tiles, but this method is more sensitive to noise. Darabi et al. proposed to use artificial neural network to segment defect images, but this algorithm is complex and has a large amount of calculation, which cannot meet the requirements of online real-time detection. The above algorithms all need to be improved in terms of detection accuracy and speed.

After searching, the Chinese patent application number 201310020370.7, the application date is January 18, 2013, and the title of the invention is: Magnetic tile surface defect feature extraction and defect classification method based on machine vision; the application first constructs a magnetic tile surface defect feature Extract the 5-scale, 8-direction Gabor filter bank, and filter the original image to obtain 40 component images; then extract the gray mean and variance features of the component images to form an 80-dimensional feature vector; and use the PCA principal component The analysis method and ICA independent component analysis method reduce the dimension of the original 80-dimensional feature vector, remove the correlation and redundancy, and obtain the 20-dimensional feature vector; normalize the preprocessing of the feature vector data, and the original data is normalized to [0, 1]; finally, the grid method and K-CV cross-validation method are used to optimize the SVM parameters, and the SVM model is trained offline with the training sample data; during online detection, the preprocessed test sample data is input to Support vector machine can realize the automatic classification and recognition of defects.

Another example is the Chinese patent application number 201110061144.4, the application date is March 14, 2011, and the title of the invention is: method and device for automatic detection of magnetic tile surface defects based on machine vision; A method for detecting surface defects of magnetic tiles. The specific detection process is: (1) place the detected magnetic tile on the conveyor belt; (2) start the CCD image acquisition device, collect the surface image of the magnetic tile and send it to the image processing unit; (3) the image processing unit passes the collected image through After image filtering, image segmentation, post-morphological processing, edge detection, etc., the processing results are transmitted to the defect detection unit; (4) the image processing results are converted into one-dimensional digital signals through feature extraction; (5) the obtained results are After training and testing by the probabilistic pattern recognition unit, the surface quality of the magnetic tiles is divided into two types: good magnetic tiles and defective magnetic tiles, so as to achieve the purpose of defect detection.

All of the above applications can filter out the interference of the surface texture of the magnetic tiles to a certain extent, and the extracted features can also reflect the defect information to a certain extent; The gray level difference of magnetic tile products is not obvious, and there is a problem of surface curvature. The detection accuracy of the surface defects of magnetic tiles is low, and further improvement is still needed.

Contents of the invention

1. The technical problem to be solved by the invention

Aiming at the problem that the existing magnetic tile surface defect detection algorithm has insufficient detection accuracy and speed, the invention provides a magnetic tile surface defect detection method based on an improved machine vision attention mechanism to realize automatic detection of magnetic tile defects. The present invention first uses morphological top-hat and bottom-hat transformations to enhance the contrast between the defect area and the background image, and then utilizes the grayscale feature comparison between each image block after the image is divided to detect whether there is a defect in the magnetic tile, effectively reducing the algorithm Calculations; Finally, in order to overcome the interference of the normal grinding texture of the magnetic tile on the defect extraction, the improved visual attention model is used to segment the image, and the defect segmentation and extraction are realized by calculating the visual salient value of the defect image block. The present invention can effectively overcome the noise interference , the algorithm has high accuracy, good adaptability to different defects, and high reliability.

2. Technical solution

In order to achieve the above object, the technical scheme provided by the invention is:

A kind of magnetic tile surface defect detection method based on the improved machine vision attention mechanism of the present invention, its steps are:

Step 1. Input the original image of the surface of the magnetic tile, and use the method of combining top-hat and bottom-hat transformation of morphology to enhance the overall gray contrast of the image;

Step 2. Divide the magnetic tile image obtained in step 1 into a*b image blocks evenly, and the side length of each image block is M*N, and then use the grayscale feature quantity of the divided image blocks to distinguish defective image blocks from non-defective Image block, to determine whether the image has defects and determine the position of the image block where the defect is located;

Step 3: Use the improved Itti visual attention mechanism model to calculate the saliency of the defect image block obtained in step 2, select the primary features to form a saliency map, normalize the fused comprehensive saliency map, and obtain the maximum attention focus is the defect area , directly segment the comprehensive saliency map;

Step 4: After obtaining the comprehensive saliency map, the Otsu threshold segmentation algorithm is used to threshold the comprehensive saliency map to extract defect areas.

Furthermore, the process of performing top hat and bottom hat transformation in step 1 is:

The calculation formulas of the top hat transformation graph T _hat (f) and the bottom hat transformation graph B _hat (f) are as follows:

Among them, f represents the original image, γ(f) represents the opening operation of the original image f, and φ(f) represents the closing operation of the original image f;

Subtract _{That (f) from the original image f to reduce the bright details of the image, and then subtract B hat (} f) to enhance the image contrast to achieve the purpose of emphasizing the defect area of the magnetic tile. The output image is represented by k ^TH , and the calculation formula is as follows :

k ^TH ＝f-λT _hat (f)-ψB _hat (f)

In the formula, a circular structure with a radius of 17mm is selected as the structural element in the morphology, and the parameter factors λ=0.1, ψ=0.1.

Further, the gray feature quantity of the image block described in step 2 includes the image block mean value W _m , the entropy value W _c superimposed on the four-direction gray-scale co-occurrence matrix and the improved image block variance W _d ; where:

The average value W _m of the image block is the local average value of the first n items after sorting the pixel gray value of the image block from small to large, and the calculation formula is as follows:

In the formula, f(x _i , y _j ) represents the pixel gray value of the coordinate (x _i , y _j ) in the local area;

The calculation formula of the entropy value Wc superimposed on the four-direction gray-level co-occurrence matrix is as follows:

In the above formula, k is the side length of the gray-scale co-occurrence matrix; p(i, j) represents the statistical probability at (i, j) in the matrix, and t represents the direction in which the gray-scale co-occurrence matrix calculates the entropy value;

The improved calculation _formula of image block variance Wd is as follows:

W _d =Std(S _N -min(S _N ))

S _N (x) = [S _N (x ₁ ), ..., S _N (x _N )]

In the above formula, Std() is the formula for finding the mean square error, S _N (x) is a 1×N matrix formed by adding the gray values in the vertical direction of the image block, and f( _xi , y _j ) represents the coordinate is the pixel gray value of point ( _xi , y _j ).

Furthermore, the criterion for judging the defective image block and the non-defective image block in step 2 is: the W _m value of the defective image block is smaller than the W _m value of the row image block, and the W _c and W _d values are greater than the row The W _c and W _d values of the image block, the W _d value of the defective image block is greater than 3 times the second largest W _d value of the image block in the row.

Furthermore, the primary features selected in step 3 include:

a. Local brightness feature: using the mean and variance of the local brightness as a reference, subtract the gray value of each pixel in the local area from the mean and variance of the gray value of the image block in the area to obtain a difference, and calculate the difference Values are processed with an exponential function to obtain a local brightness saliency map S _l (x, y), and the calculation formula is as follows:

I _local (x _i , y _j )=f(x _i , y _j )-(m _local (x, y)+μ·d _local (x, y))

In the formula, m _local (x, y) represents the local gray value average, d _local (x, y) represents the local gray value variance, f(x _i , y _j ) represents the gray value, μ is the variance control factor, take The value range is 0~1;

b. Global brightness feature: use the global average as a reference, subtract the gray value of each pixel from it, and perform exponential function processing on the difference to obtain the global brightness significant S _g (x, y), the calculation formula as follows:

I _global (x _i , y _i )＝|f(x _i , y _i )-m _global |

In the formula, m _global is the gray average value of the entire magnetic tile image, and sum(I _global ( _xi , y _i )) is the sum of all I _global ( _xi , y _i ) values;

c. Frequency feature: use the DoG filter to obtain the frequency feature saliency map S _G (x, y), the calculation formula is as follows:

S _G (x, y)＝|m _global -I _G (x, y)|

In the formula, I _G (x, y) is the DoG filtered image, σ ₁ and σ ₂ are the standard deviation of the Gaussian kernel function, σ ₁ :σ ₂ =5:1.

Furthermore, said σ ₁ =0.5, σ ₂ =0.1.

Furthermore, in step 3, the obtained local brightness saliency map, global brightness saliency map, and frequency feature saliency map are fused, and the calculation formula for generating a comprehensive saliency map is as follows:

3. Beneficial effect

Compared with the existing known technology, the technical solution provided by the invention has the following remarkable effects:

(1) A magnetic tile surface defect detection method based on an improved machine vision attention mechanism of the present invention enhances the contrast between the defect region and the background image in the image through morphological top-hat and bottom-hat transformations, and suppresses the grayscale of the high-brightness region Value, so that the overall gray contrast of the image is enhanced, and it is convenient to identify the defect area;

(2) A magnetic tile surface defect detection method based on the improved machine vision attention mechanism of the present invention, combined with the idea of image segmentation, utilizes the difference between the gray feature value of the defective image block and the non-defective image block to effectively distinguish the defect Image blocks and non-defective image blocks, to determine whether there is a defect in the image and determine the position of the image block where the defect is located, which improves the efficiency of the algorithm, reduces the calculation time and reduces the complexity of the segmentation algorithm;

(3) A magnetic tile surface defect detection method based on the improved machine vision attention mechanism of the present invention improves the visual attention mechanism model, and realizes defect segmentation and extraction by calculating the visual saliency value of the defect image block, which can effectively overcome the magnetic tile surface defect detection method. Texture interference, accurately complete the defect extraction on the surface of the magnetic tile;

(4) A magnetic tile surface defect detection method based on an improved machine vision attention mechanism of the present invention effectively overcomes problems such as uneven brightness, small defect area of magnetic tiles, and texture interference of the magnetic tile itself, and can be quickly and effectively It can accurately extract all kinds of magnetic tile defects and has strong adaptability.

Description of drawings

Fig. 1 is the flow chart of magnetic tile surface defect detection algorithm of the present invention;

(a)～(h) among Fig. 2 is original defect figure of the present invention;

(a)-(h) in Fig. 3 are saliency map fusion maps of the present invention;

(a)-(h) in FIG. 4 are image segmentation diagrams of the present invention.

Detailed ways

In order to further understand the content of the present invention, the present invention will be described in detail in conjunction with the accompanying drawings and embodiments.

Example 1

In conjunction with FIG. 1, a method for detecting surface defects of magnetic tiles based on an improved machine vision attention mechanism in this embodiment includes the following steps:

Step 1: Input the original image of the surface of the magnetic tile. During image processing, some areas with low gray values are easily confused with defect areas, which interferes with the processing results. This embodiment proposes a method of combining the top-hat and bottom-hat transformations of morphology to improve the contrast between the defect area and the background, suppress the gray value of the high-brightness area, enhance the overall gray contrast of the image, and facilitate the identification of the defect area .

The process of performing top-hat and bottom-hat transformations is:

The top hat transformation map T _hat (f) is the opening operation γ(f) of the image subtracted from the original image f, while the bottom hat transformation map B _hat (f) is the closing operation φ(f) of the original image minus f, Calculated as follows:

After the top hat transformation, the highlighted area shows the high gray value area of the original image f, and after the bottom hat transformation, the highlighted area shows the low gray value area of the original image f. Therefore, in order to further reduce the gray value of the brighter part, you can subtract _{That (f) from the original image f to reduce the bright details of the image, and then subtract B hat (} f) to enhance the image contrast and realize the emphasis on magnetic tile defects area. The output image is represented by k ^TH , and the calculation formula is as follows:

k ^TH ＝f-λT _hat (f)-ψB _hat (f)

In the above formula, a circular structure with a radius of 17mm is selected as the structural element in the morphology, and the parameter factors are λ=0.1 and ψ=0.1.

Step 2: In order to improve the efficiency of the algorithm, reduce the calculation time and reduce the complexity of the segmentation algorithm, it is necessary to quickly detect whether there is a defect area in an image before image segmentation. In this embodiment, firstly, combining the idea of image segmentation, the magnetic tile image obtained in step 1 is evenly divided into a*b image blocks, and the side length of each image block is M*N. The values of area, M, and N need multiple tests to be determined. Then, using the grayscale statistics of the divided image blocks, that is, the mean value of the divided image blocks, the entropy value of the four-direction gray-scale co-occurrence matrix superimposition and the improved image block variance and other three gray-scale feature quantities, by comparing The grayscale feature quantities of different image blocks of the input image can effectively distinguish defective image blocks from non-defective image blocks, determine whether the image has defects and determine the position of the image block where the defect is located.

Using the above three kinds of gray feature quantities, the specific process of distinguishing defective image blocks from non-defective image blocks is as follows:

Since the image is divided into a row and b column, it is necessary to calculate the three grayscale feature quantities of a*b image blocks, specifically:

a. Mean value W _m : Since the defect area is relatively small compared to the area of the entire image block, it contributes little to the image mean value. Therefore, when calculating the mean value, the pixels involved in the calculation of the mean value can be counted and the statistical range is limited to the defect area as much as possible. In this embodiment, the pixel gray values are sorted from small to large, and then the local mean W _m of the first n items is counted. In practice, gray histograms can be used to achieve this. The value of n is 250, which is more appropriate. Calculated as follows:

In the formula, f(x _i , y _j ) represents the pixel gray value of the coordinate (x _i , y _j ) in the local area;

b. Entropy value W _c : The gray level co-occurrence matrix mainly reflects the gray level distribution information, and the elements in the co-occurrence matrix are composed of the number of pixel point pairs with the same gray value separated by a certain distance. In order to reduce the amount of computation, this embodiment compresses the gray level of the image block from 0-255 to 0-15 and then uses some scalars to represent the gray-level co-occurrence matrix, including energy, entropy, moment of inertia and correlation. The entropy value W _g in the gray level co-occurrence matrix can reflect the complexity of the image. In order to increase the significance of the entropy value, the entropy values in the four directions can be added together, and the sum is W _c . Calculated as follows:

In the above formula, k is the side length of the gray level co-occurrence matrix. When the gray level is compressed to 16, k=16; p(i, j) represents the statistical probability at (i, j) in the matrix, and t represents the gray The degree co-occurrence matrix calculates the direction of the entropy value.

c. Variance W _d : In order to reduce the influence of the magnetic tile texture, this embodiment first adds the gray values in the vertical direction in the image block to form a 1×N matrix S _M (y); Addition affects the role of the defect area in the calculation process. When calculating, the minimum value in the matrix is subtracted from the matrix items, and finally the mean square error of the matrix is calculated. The calculation formula is as follows:

W _d =Std(S _N -min(S _N ))

S _N (x) = [S _N (x ₁ ), ..., S _N (x _N )]

In the above formula, Std() is the formula for finding the mean square error, and S _N (x) is a 1×N matrix f(x _i , y _j ) formed by adding the gray values in the vertical direction of the image block, indicating that the coordinates are (x _i , y _j ) pixel gray value.

Count the mean value, entropy value and variance of each image block after each row is divided into blocks. The criterion for judging whether there is a defective block is: the _Wm value of the defective image block is smaller than the image block in the row, and the _Wc and _Wd are larger than the image block in the row. It is worth noting that the _Wd value of the defective image block is is much larger than the non-defective image block, so a certain multiple relationship should be set when W _d is used as the criterion. In this embodiment, the multiple is taken to be that the largest W _d value in a row of image blocks is greater than 3 times the second largest W _d . If the parameters of an image block in a row of image block statistics table meet the above conditions, it is judged that the image block is a defective image block.

Step 3: Since the surface texture of the magnetic tile is relatively rich, it is difficult to directly perform threshold segmentation. In this embodiment, the improved Itti visual attention mechanism model is used to calculate the saliency of the defect image block obtained in step 2. From the perspective of global saliency value and local saliency value, consider re-selecting the primary features to form a saliency map, normalize the fused comprehensive saliency map to [0, 1], and then obtain the maximum focus of attention (that is, the position with a saliency value of 1 ) is the defect area. Since the block size of the defect image is small, there is no need to consider other focus points, so the integrated saliency map can be directly segmented. The specific process is:

(1) Primary feature selection

a. Local brightness feature: using the mean and variance of the local brightness as a reference, subtract the gray value of each pixel in the local area from the mean and variance of the gray value of the image block in the area to obtain a difference, and calculate the difference Values are processed with an exponential function to obtain a local brightness saliency map S _l (x, y), and the calculation formula is as follows:

I _local (x _i , y _j )=f(x _i , y _j )-(m _local (x, y)+μ·d _local (x, y))

In the formula, m _local (x, y) represents the local gray value average, d _local (x, y) represents the local gray value variance, f(x _i , y _j ) represents the gray value, and μ is the value range of the variance control factor It is 0~1.

b. Global brightness feature, in the large defect area, the middle part of the defect area lacks sufficient local contrast, so the global average value is used as a reference, and the gray value of each pixel is subtracted from it, and the difference value is calculated processed to obtain a global brightness saliency map. The calculation formula of the global brightness saliency map S _g (x, y) is as follows:

I _global (x _i , y _j )=|f(x _i , y _j )-m _global |

In the formula, m _global is the gray average value of the entire magnetic tile image, and sum(I _global ( _xi , y _j )) is the sum of all I _global ( _xi , y _j ) values.

c. Frequency characteristics: Since the DoG filter can enhance the visibility of edges and other details while filtering out interference, the DoG filter is selected, and the calculation formula is as follows:

In the formula, σ ₁ and σ ₂ are the standard deviation of the Gaussian kernel function, and their proportional relationship controls the bandwidth of the filter. When σ ₁ = 1.6σ ₂ , it is an edge detector. When σ ₁ is infinite, the DC component Filter out, that is, the background image is filtered out.

When processing image data, a smaller Gaussian kernel window is used to filter out texture and other interference information, generally [1,4,6,4,1]/16 is used. Using the DoG filter, the frequency feature saliency map is obtained by the following formula:

S _G (x, y)＝|m _global -I _G (x, y)|

In the formula, S _G (x, y) is the saliency map of frequency features, and I _G (x, y) is the image filtered by DoG. In this embodiment, σ ₁ :σ ₂ =5:1, σ ₁ =0.5, σ ₂ =0.1.

(2) Feature fusion: Since the local brightness saliency map, the global brightness saliency map, and the frequency feature saliency map have certain limitations in expressing the characteristics of specific directions, a comprehensive saliency map is generated by fusion for image segmentation. Calculated as follows:

Step 4: After obtaining the comprehensive saliency map, select the Ostu threshold segmentation algorithm to threshold the comprehensive saliency map, and obtain the salient area and non-salient area (foreground and background) of the comprehensive saliency map, which can realize the salient area of the original image (defective region) segmentation. The segmented salient area does not completely correspond to the defect area of the original image, and there may be edge areas and isolated salient points, but these details can be eliminated by performing morphological opening on the thresholded image.

(a)～(h) in Fig. 2 is the original defect map of the present invention, and (a)～(h) in Fig. 3 is the saliency map fusion map of the present invention, it can be seen from Fig. 3 that relative to primary features The saliency map, the total saliency map has a significant improvement in suppressing texture information. (a)-(h) in Fig. 4 are the image segmentation effect diagrams of the present invention, and it can be seen from Fig. 4 that after the comprehensive salient image is segmented, the defect area is basically extracted.

The above schematically describes the present invention and its implementation, which is not restrictive, and what is shown in the drawings is only one of the implementations of the present invention, and the actual structure is not limited thereto. Therefore, if a person of ordinary skill in the art is inspired by it, without departing from the inventive concept of the present invention, without creatively designing a structural mode and embodiment similar to the technical solution, it shall all belong to the protection scope of the present invention .

Claims (6)

1. a kind of magnetic tile surface defect detection method based on improved machine vision attention mechanism, step are：

Step 1: input magnetic shoe surface original image, converts the method combined, enhancing figure using morphologic top cap and bottom cap As overall intensity contrast degree；

Step 2: magnetic shoe image uniform obtained by step 1 is divided into a*b image block, each image block side length is M*N, then sharp Defect image block and non-defective image block are distinguished with the gray feature amount of the image block after piecemeal, judges image with the presence or absence of defect And determine tile location where defect；

Step 3: using the significance for improving defect image block obtained by Itti vision noticing mechanism model calculating step 2, selection Primary features are normalized fused comprehensive notable figure to form notable figure, and obtaining maximum focus-of-attention is to lack Fall into region；Wherein, the primary features of selection include：

A. local luminance feature：Using the mean value and variance of local luminance as reference, with each pixel in regional area The mean value and variance of gray value and the image block gray value in the region subtract each other to obtain difference, and carry out at exponential function to the difference Reason, obtains local luminance notable figure S_l(x, y), calculation formula are as follows：

I_local(x_i, y_j)=f (x_i, y_j)-(m_local(x, y)+μ d_local(x, y))

In formula, m_local(x, y) indicates local gray level average value, d_local(x, y) indicates local gray level variance, f (x_i, y_j) indicate to sit It is designated as (x_i, y_j) point grey scale pixel value, μ be variance controlling elements, value range be 0~1；

B. global brightness：Using global average value as reference, subtracted each other therewith with the gray value of each pixel, and to difference Value carries out exponential function processing, obtains global brightness notable figure S_g(x, y), calculation formula are as follows：

I_global(x_i, y_j)=| f (x_i, y_j)-m_global|

In formula, m_globalFor the average gray of whole picture magnetic shoe image, sum (I_global(x_i, y_j)) it is all I_global(x_i, y_j) value Sum；

C. frequecy characteristic：Using DoG filter, frequecy characteristic notable figure S is obtained_G(x, y), calculation formula are as follows：

S_G(x, y)=| m_global-I_G(x, y) |

In formula, I_G(x, y) is through the filtered image of DoG, σ₁And σ₂It is the standard deviation of gaussian kernel function, σ₁∶σ₂=5: 1；

Step 4: selecting Otsu threshold partitioning algorithm to comprehensive notable figure thresholding after obtaining comprehensive notable figure, defect is extracted Region.

2. a kind of magnetic tile surface defect detection side based on improved machine vision attention mechanism according to claim 1 Method, it is characterised in that：Step 1 carries out top cap and the process of bottom cap transformation is：

Top cap Transformation Graphs T_hat(f) and bottom cap Transformation Graphs B_hat(f) calculation formula is as follows：

Wherein, f indicates that original image, γ (f) indicate that the opening operation of original image f, φ (f) indicate the closed operation of original image f；

Original image f is subtracted into T_hat(f) it reduces the bright detail of image, then subtracts B_hat(f) enhance picture contrast, it is real It now emphasizes the purpose of magnetic shoe defect area, exports image k^THIt indicates, calculation formula is as follows：

k^TH=f- λ T_hat(f)-ψB_hat(f)

In formula, select the circular configuration that radius is 17mm as the structural element in morphology, parameter factors λ=0.1, ψ= 0.1。

3. a kind of magnetic tile surface defect detection based on improved machine vision attention mechanism according to claim 1 or 2 Method, it is characterised in that：The gray feature amount of image block described in step 2 includes image block mean value W_m, it is total to four direction gray scales The entropy W of raw matrix superposition_cWith improved image block variance W_d；Wherein：

Image block mean value W_mFor n before counting after image block grey scale pixel value sorts from small to large local mean values, calculation formula It is as follows：

In formula, f (x_i, y_j) indicate that coordinate is (x in regional area_i, y_j) point grey scale pixel value；

To the entropy W of four direction gray level co-occurrence matrixes superposition_cCalculation formula is as follows：

In above formula, k is the side length of gray level co-occurrence matrixes；P (i, j) indicates the statistical probability in matrix at (i, j), and t indicates ash Spend the direction that co-occurrence matrix calculates entropy；

Improved image block variance W_dCalculation formula is as follows：

W_d=Std (S_N(x)-min(S_N(x)))

S_N(x)=[S_N(x₁) ..., S_N(x_N)]

In above formula, Std () is to ask mean square deviation formula, S_N(x) for the gray value of vertical direction in image block is added to be formed 1 × The matrix of N, f (x_i, y_j) indicates coordinate be (x_i, y_j) point grey scale pixel value.

4. a kind of magnetic tile surface defect detection side based on improved machine vision attention mechanism according to claim 3 Method, it is characterised in that：The standard for judging defect image block and non-defective image block in step 2 is：The W of defect image block_mValue is equal Less than the W of locating row image block_mValue, W_cValue is greater than the W of locating row image block_cValue, the W of defect image block_dValue is greater than locating row figure As the second largest W of block_d3 times of value.

5. a kind of magnetic tile surface defect detection side based on improved machine vision attention mechanism according to claim 4 Method, it is characterised in that：The σ₁=0.5, σ₂=0.1.

6. a kind of magnetic tile surface defect detection side based on improved machine vision attention mechanism according to claim 5 Method, it is characterised in that：Local luminance notable figure, global brightness notable figure, frequecy characteristic notable figure in step 3 for obtaining It is merged, the calculation formula for generating comprehensive notable figure is as follows：