CN103871053A - Vision conspicuousness-based cloth flaw detection method - Google Patents

Abstract

The invention discloses a cloth defect detection method based on visual salience, which comprises the following steps: (1) collecting images; (2) processing brightness features; (3) processing direction features; (4) multi-channel superposition and normalizing processing; (5) Grayscale image processing; (6) Binarization processing; (7) Defect area judgment. Compared with the traditional cloth defect detection method, the present invention reduces the complexity of calculation, improves the recognition rate, and can accurately locate , and at the same time, it avoids false detections easily caused by the case where the gray value of the salient image of the detected cloth intact image is higher than the gray value of the intact part of the defect image, effectively reducing the interference of the background during the detection process, and reducing It prevents the misjudgment of the target area obtained by the adaptive threshold segmentation of the image of the intact cloth as the defective area.

Description

A method of cloth defect detection based on visual saliency

technical field

The invention relates to a cloth defect detection method, in particular to a cloth defect detection method based on visual salience.

Background technique

In modern textile production, quality control and testing are very important, and cloth defect detection is a particularly critical component. At present, domestic textile enterprises mainly use manual testing methods, and the detection speed of human eyes is limited, and the detection results are easily affected. Influenced by subjective factors, false detection and missed detection are prone to occur. Replacing manual cloth defect detection with advanced automatic detection technology is an important measure to improve detection efficiency, reduce labor, reduce labor intensity and ensure cloth quality. Scholars at home and abroad have made many outstanding achievements in the research of automatic detection methods.

The cloth defect detection algorithm mainly judges the defect according to the detected cloth texture and shape, and there are roughly three types: statistical method, spectral analysis method and model-based method. The most common of these is the spectrum analysis method using Fourier transform, wavelet transform and Gabor analysis. The Fourier transform is a global transformation of the image, so it cannot accurately locate the defect; the transformation detection performance of the Gabor analysis is better, but it needs to perform two-dimensional filtering and fusion on the multi-channel direction, which greatly increases the computational complexity; the wavelet transform has a good It has the advantages of local time-frequency analysis and fast calculation speed, but its poor direction selectivity makes it unable to describe the characteristics of two-dimensional graphs well, resulting in unsatisfactory detection results.

Humans can quickly and effectively identify defects, regardless of the intensity of reflection, changes in the shape of defects, differences caused by different production processes, and so on. When the cloth is presented to the human eye, its flaws are more obvious and can attract visual attention. That is to say, the information to be discovered in the graph is often concentrated in some key areas, which are usually called salient areas. If these salient regions, i.e., visually attractive regions, can be accurately extracted, we can effectively find the needed graph information. To extract salient regions from graphs, computational models based on visual attention have been proposed. In these models, the saliency of visual regions is represented by a saliency map. A saliency map is a two-dimensional map, and the value of each pixel in it represents the saliency of the corresponding pixel in the original image. The present invention is based on the principle of visual salience, and provides a method for detecting defects on a cloth pattern.

Contents of the invention

The purpose of the present invention is to provide a cloth defect detection method based on visual saliency, which simulates the human visual attention mechanism, based on the GBVS (Graph-based Visual Saliency) model, by extracting the brightness and direction features of the cloth to generate visual The saliency map, by analyzing the gray level distribution relationship of the visual saliency map, judges whether the energy of the visual saliency map is concentrated, and then judges whether there is a defect in the cloth.

In order to achieve the above-mentioned purpose of the invention, the technical solution adopted in the present invention is: a method for detecting cloth defects based on visual salience, comprising the following steps:

(1) Collect images, collect images of cloth, and obtain initial grayscale images;

(2), the GBVS model generates a salient image, and generates a comprehensive salient map of the initial grayscale image according to the GBVS model;

(3) Grayscale image processing, reducing the grayscale value of each pixel in the comprehensive saliency map, so as to obtain the final saliency map after weakening the background saliency;

(4) Binarization processing, performing adaptive threshold segmentation on the final saliency map according to the maximum inter-class variance method, to obtain a binary map with prominent and salient target parts;

(5) Defect area judgment, by comparing the gray sum of squares of the corresponding target area in the final saliency map with the gray sum of squares of the entire area of the integrated saliency map to judge whether the target area is a defect.

In a further technical solution, step (2) includes the following sub-steps:

(a) Luminance feature processing, performing Gaussian pyramid filtering on the initial grayscale image to obtain a set of luminance feature filtering results at different scales, and for each of the luminance feature filtering results, according to the difference between the pixels and the Ou The distance to establish their own Markov chains, and then find their Markov balance

distribution, performing normalization processing on the set of luminance feature filtering results at different scales after the Markov equilibrium distribution, to obtain a luminance feature saliency map;

(b), directional feature processing, performing Gabor pyramid filtering on the initial grayscale image in four directions, correspondingly obtaining the directional feature filtering results under four groups of different scales, for each of the directional feature filtering results, according to its The difference between pixels and the Euclidean distance establish their respective Markov chains, and then find their Markov equilibrium distribution, and normalize the directional feature filtering results of the set of different scales after the Markov equilibrium distribution , to obtain a directional feature saliency map;

(c), multi-channel superposition normalization processing, linearly adding the luminance feature saliency map in step (a) and the direction feature saliency map in step (b) and performing normalization processing to obtain comprehensive saliency picture;

Wherein, the steps (a) and (b) are in no sequence.

In a further technical solution, step (3) adopts the following method: calculate the average gray value of each pixel in the comprehensive saliency map, and judge whether the gray value of each pixel is greater than the average gray value, if the If the gray value is less than the average gray value, the gray value is reset to zero; otherwise, the gray value is equal to the difference between the gray value and the average gray value, and the final result of weakening the background salience is obtained. Significant figure.

In a further technical solution, step (5) adopts the following method: calibrate the area corresponding to the target part in the final saliency map as the target area, and calculate the gray square sum of the pixels of the target area in the final saliency map and the obtained The sum of the gray squares of the pixels in the entire area of the comprehensive saliency map, and then calculate the ratio of the gray square sum of the target area to the gray square sum of the entire area, if the ratio is lower than the judgment threshold, the The target area is a blemish area, otherwise the target area is not a blemish area.

In a further technical solution, the value range of the judgment threshold is 0.15-0.4.

Due to the use of the above-mentioned technical solutions, the present invention has the following advantages compared with the prior art:

1. By analyzing the visual salience of the input cloth image, it is judged whether a binary image with a prominent target area is obtained, and then whether there is a defect in the cloth is judged. Compared with the traditional cloth defect detection method, the present invention reduces the calculation cost. Complexity, improved recognition rate, and accurate positioning;

2. By reducing the gray value of the points whose pixel gray value is higher than the average value in the comprehensive saliency map, the gray value of the point whose pixel gray value is lower than the average value in the comprehensive saliency map is zeroed to weaken the background (regarding the intact area of the cloth as the background) saliency, so that there is an obvious gap between the saliency of the flawed area and the intact area, avoiding the gray value ratio of the detected intact area of the cloth due to the generation principle of the comprehensive saliency map It is easy to cause false detection when the gray value of the intact part in the defect area is even higher, which effectively reduces the interference of the background during the detection process;

3. Since the human eye always has a relatively concentrated area when observing the uniform texture map, the detection method of cloth defects based on visual salience is easy to misjudge the target part obtained by segmenting the intact cloth image as a defect area. By extracting The image feature information of the target area is used to judge the energy, judge whether the target area is an energy concentrated area, finally determine the defect area, and calculate the gray square sum of the pixel points corresponding to the target part in the final saliency map and the resulting The square sum of pixel gray levels in the entire area of the comprehensive saliency map is calculated, and then the ratio between the two is calculated, and the target area is judged whether the energy of the target area is relatively concentrated by the ratio. If the ratio is greater than the judgment threshold, that is, the energy is relatively concentrated, the target The region is a defect, otherwise the target region is not a defect. Through the above technical solution, the occurrence of misjudgment is reduced.

Description of drawings

Fig. 1 is the implementation flowchart of the method for detecting cloth defects based on visual salience disclosed by the present invention;

Fig. 2 is a step-by-step detection result figure of the cloth intact figure disclosed by the present invention;

Fig. 3 is a step-by-step detection result diagram of a typical cloth defect diagram disclosed by the present invention;

Fig. 4, Fig. 5, Fig. 6 and Fig. 7 are diagrams of detection results of other cloth defect diagrams disclosed by the present invention.

Detailed ways

The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

Embodiment: referring to Fig. 1, as shown in the legend wherein, above-mentioned cloth flaw detection method comprises the following steps:

(1) Collect images, collect images of cloth through industrial cameras, and obtain initial grayscale images , Figure 2(a) is the initial image of the cloth intact image, and Figure 3(a) is the initial grayscale image of the typical cloth defect image.

(2), Brightness feature processing:

进行高斯金字塔滤波，金字塔滤波是指对初始灰度图不断进行1/2降采样和滤波，滤波的尺度因子随图的减小而减小，得到一组不同尺度下的滤波结果，本例中金字塔等级为2，即得到两个尺度下的不同亮度特征滤波结果，即两张亮度特征图

； a. The above initial grayscale image input through a two-dimensional Gaussian filter

Perform Gaussian pyramid filtering. Pyramid filtering refers to continuous 1/2 downsampling and filtering of the initial grayscale image. The scale factor of the filter decreases as the image decreases, and a set of filtering results at different scales is obtained. In this example The pyramid level is 2, that is, the filtering results of different brightness features at two scales are obtained, that is, two brightness feature maps

;

是初始灰度图，

是对亮度特征在尺度因子

下用高斯函数滤波得到的高斯金字塔，

为二维高斯滤波器的高斯函数，

为相应图中像素点的位置，

为相应图的尺度因子； in,

is the initial grayscale image,

is the scale factor for brightness features in

The Gaussian pyramid obtained by filtering with the Gaussian function,

is the Gaussian function of the two-dimensional Gaussian filter,

is the position of the pixel in the corresponding image,

is the scale factor of the corresponding graph;

进行马尔科夫平衡分布，进而归一化处理，得到亮度特征显著图； b. For each of the brightness feature maps

Carry out Markov equilibrium distribution, and then normalize to obtain the brightness feature salience map ;

为

，任意两像素点之间的距离为

，则任意两像素点

和

距离： Let each brightness feature map

for

, the distance between any two pixels is

, then any two pixels

and

distance:

每两个像素点都两两连接，每条边代表两点间的权重，点

到点

的权重

为: feature map

Every two pixels are connected in pairs, and each edge represents the weight between two points.

to the point

the weight of

for:

in is the Gaussian weighting parameter;

That is to say, the weight between any two pixels is determined by the gray level difference between the two pixels and the distance between the points, and the weight in the opposite direction is the same. The above process is to find the maximum eigenvalue corresponding to the weight matrix corresponding to each brightness feature map The process of feature vector, that is, the process of finding the equilibrium state of each brightness feature map;

进行归一化处理，得到一张亮度特征显著图

。 Balance the distribution of two brightness features

Perform normalization processing to obtain a luminance feature salience map

.

(3), direction feature processing:

； a. The initial grayscale image of the input through the Gabor filter Perform Gabor pyramid filtering in four directions. Pyramid filtering refers to continuous 1/2 downsampling and filtering of the initial grayscale image. The scale factor of the filter decreases with the decrease of the image, and four groups of filters at different scales are obtained. As a result, the pyramid level in this example is 2, that is, four groups of different directional feature filtering results at two scales are obtained, that is, eight directional feature maps

;

为Gabor滤波器的Gabor函数，

为相应图中像素点的位置，为尺度因子, 为方向因子（取

），

为Gabor滤波器的正弦波频率，； in, It is the Gabor pyramid obtained by filtering the directional feature at the scale with the Gabor function of the direction,

is the Gabor function of the Gabor filter,

is the position of the pixel in the corresponding image, is the scale factor, is the direction factor (take

),

is the sine wave frequency of the Gabor filter,;

进行马尔科夫平衡分布，进而归一化处理，得到亮度特征显著图

， b. For each direction feature map

Carry out Markov equilibrium distribution, and then normalize to obtain the brightness feature salience map

,

为

，任意两像素点之间的距离为

，则任意两像素点

和

距离： Let each orientation feature map

for

, the distance between any two pixels is

, then any two pixels

and

distance:

到点

的权重为: feature map Every two pixels are connected in pairs, and each edge represents the weight between two points.

to the point

the weight of for:

in is the Gaussian weighting parameter;

That is to say, the weight between any two pixels is determined by the gray difference between the two pixels and the distance between the points, and the weight in the opposite direction is the same. The above process is to find the maximum eigenvalue corresponding to the weight matrix corresponding to each direction feature map. The process of eigenvectors is the process of finding the equilibrium state of each directional feature map;

进行归一化处理，得到方向特征显著图

； Balance the distribution of eight directional features

Perform normalization processing to obtain a directional feature saliency map

;

线性相加并进行归一化处理，得到与上述初始灰度图

大小一致的综合显著图

，如图2（b）为布匹完好图的综合显著图，如图3（b）为布匹典型瑕疵图的综合显著图。 (4) Superimposed and normalized processing, the above-mentioned brightness feature saliency map and directional feature saliency map

Add linearly and perform normalization processing to obtain the above initial grayscale image

Synthesized saliency map with consistent size

, as shown in Figure 2(b) is the comprehensive saliency map of the cloth intact map, and Figure 3(b) is the comprehensive saliency map of the typical cloth defect map.

的灰度平均值，比较上述综合显著图各像素点灰度值与灰度平均值，如果上述像素点灰度值小于灰度平均值，则对上述像素点灰度值进行归零处理，否则上述像素点灰度值等于其与灰度平均值的差值，得到减弱背景灰度的最终显著图

，如图2（c）为布匹完好图的最终显著图，如图3（c）为布匹典型瑕疵图的最终显著图。 (5) Grayscale image processing, calculating the above comprehensive saliency map

Compared with the above comprehensive saliency map The gray value of each pixel and the average value of the gray value. If the gray value of the above pixel is smaller than the average value of the gray value, the gray value of the above pixel is zeroed, otherwise the gray value of the above pixel is equal to its gray value The difference between the mean values to obtain the final saliency map with reduced background grayscale

, Figure 2(c) is the final saliency map of the cloth intact map, and Figure 3(c) is the final saliency map of the typical cloth defect map.

进行自适应阈值分割，得到具有突出显著的目标部分的二值图

，如图2（d）为布匹完好图的二值图，如图3（d）为布匹典型瑕疵图的二值图。 (6), binarization processing, according to the maximum between-class variance method, the above-mentioned final saliency map

Perform adaptive threshold segmentation to obtain a binary image with prominent target parts

, as shown in Figure 2(d) is the binary image of the good cloth image, and Figure 3(d) is the binary image of the typical defect image of the cloth.

中对应于所述目标部分的区域为目标区域，计算上述最终显著图

的目标区域的像素点灰度平方和及所述综合显著图

的整个区域的像素点灰度平方和，进而计算两者的比值。在本发明方法中判断阈值取值范围为0.15～0.4，本例判断阈值取为0.2，如果上述比值低于判断阈值0.2时，则上述目标区域为瑕疵区域，否则上述目标区域不是瑕疵区域，如图2（e）为布匹完好图的判断结果图，如图3（e）为布匹典型瑕疵图的判断结果图。 (7) Defect area judgment, calibration of the above final saliency map

The area corresponding to the target part in is the target area, and the above-mentioned final saliency map is calculated

The gray sum of squares of pixels in the target area and the integrated saliency map

The sum of the squares of the pixel gray levels of the entire area, and then calculate the ratio of the two. In the method of the present invention, the judgment threshold value ranges from 0.15 to 0.4, and the judgment threshold value in this example is 0.2. If the above-mentioned ratio is lower than the judgment threshold value of 0.2, the above-mentioned target area is a flawed area, otherwise the above-mentioned target area is not a flawed area, such as Figure 2(e) is the judgment result diagram of the cloth intact image, and Figure 3(e) is the judgment result diagram of the cloth typical defect image.

Calculate the ratio of the gray square sum of the segmented area to the entire area, and judge whether the energy of the target area is relatively concentrated by the ratio. If the ratio is greater than the judgment threshold, that is, the energy is relatively concentrated, it is a defect; otherwise, it is not a defect.

See Figures 4-7, which are the detection results of several other typical defect images of cloth, where (a) is the initial image, (b) is the comprehensive saliency image, and (c) is the judgment result image.

Claims (5)

1. A method for detecting cloth defects based on visual salience, comprising the steps of: (1) Collect images, collect images of cloth, and obtain initial grayscale images; (2), the GBVS model generates a salient image, and generates a comprehensive salient map of the initial grayscale image according to the GBVS model; (3) Grayscale image processing, reducing the grayscale value of each pixel in the comprehensive saliency map, so as to obtain the final saliency map after weakening the background saliency; (4) Binarization processing, performing adaptive threshold segmentation on the final saliency map according to the maximum inter-class variance method, to obtain a binary map with prominent and salient target parts; (5) Defect area judgment, by comparing the gray sum of squares of the corresponding target area in the final saliency map with the gray sum of squares of the entire area of the integrated saliency map to judge whether the target area is a defect. 2. The method for detecting cloth defects based on visual salience according to claim 1, wherein step (2) includes the following sub-steps: (a) Luminance feature processing, performing Gaussian pyramid filtering on the initial grayscale image to obtain a set of luminance feature filtering results at different scales, and for each of the luminance feature filtering results, according to the difference between the pixels and the Ou The distance to establish their own Markov chains, and then find their Markov balance distribution, performing normalization processing on the set of luminance feature filtering results at different scales after the Markov equilibrium distribution, to obtain a luminance feature saliency map; (b), directional feature processing, performing Gabor pyramid filtering on the initial grayscale image in four directions, correspondingly obtaining the directional feature filtering results under four groups of different scales, for each of the directional feature filtering results, according to its The difference between pixels and the Euclidean distance establish their respective Markov chains, and then find their Markov equilibrium distribution, and normalize the directional feature filtering results of the set of different scales after the Markov equilibrium distribution , to obtain a directional feature saliency map; (c), multi-channel superposition normalization processing, linearly adding the luminance feature saliency map in step (a) and the direction feature saliency map in step (b) and performing normalization processing to obtain comprehensive saliency picture; Wherein, the steps (a) and (b) are in no sequence. 3. According to claim 1, the cloth defect detection method based on visual saliency is characterized in that the step (3) adopts the following method: calculating the average gray level of each pixel in the comprehensive saliency map, and judging the Whether the grayscale value of the pixel is greater than the grayscale average value, if the grayscale value is smaller than the grayscale average value, the grayscale value is zeroed, otherwise the grayscale value is equal to its The difference of the mean value of the gray level is used to obtain the final saliency map which weakens the saliency of the background. 4. According to claim 1, the cloth defect detection method based on visual saliency is characterized in that step (5) adopts the following method: calibrate the area corresponding to the target part in the final saliency map as the target area, and calculate respectively The sum of squares of pixel grayscales in the target area of the final saliency map and the sum of squares of grayscales of pixels in the entire area of the integrated saliency map, and then calculate the sum of squares of grayscales in the target area and the square sum of grayscales in the entire area and the ratio, if the ratio is lower than the judgment threshold, the target area is a flawed area; otherwise, the target area is not a flawed area. 5 . The cloth defect detection method according to claim 4 , wherein the value range of the judgment threshold is 0.15˜0.4.

Images