With the rapid development of visual inspection technology, computer technology, and image processing technology, machine vision technology has become more and more mature, and the role of quality inspection and control in the steel industry is becoming more and more obvious and important. Defects on the surface of the strip are a key factor affecting the quality inspection process. Its inspection plays an extremely important role in improving the final quality. For a long time, traditional manual inspection methods cannot meet actual production needs, so in-depth research on steel surface defect inspection systems has become the consensus of today’s steel companies. The accuracy and low performance of traditional detection methods can no longer meet the needs of people and society. The surface defect detection method based on machine vision has the characteristics of high accuracy, fast processing speed, and intelligent processing, which is the main trend of surface defect detection. We select a steel plate; take the invariant moment features of the cracks, holes, scratches, oil stains, and other images on it; extract the data results; and analyze them. Then, we read the texture features of these defect images again, extract the data results, and analyze them. The experimental results prove that after the mean value filter and Gaussian filter process the image, the mean variance value MSE is relatively large (
Ophthalmic diseases have more complex causes, more different diseases, and more variable conditions and belong to clinical ophthalmic diseases [
Defects on the surface of the strip are a key indicator for evaluating the quality of the film. The effective method of detecting defects has attracted people’s attention for a long time. Researchers at home and abroad have done a lot of work for this and achieved some results. Ohkubo et al. proposed a detection system that uses a laser as a scanning light source, a 12-sided reflective prism and a cylindrical mirror as an optical system, and a photomultiplier tube to receive the detection system [
The existing extraction methods are mainly aimed at normal retinal images, which are not universal; this paper adopts an improved maximum classification algorithm based on constant torque to solve the inefficient ground-to-ground detection problem, using gray scale and texture feature extraction methods, feature selection methods based on principal component analysis, and defect classification algorithms based on support carrier machines. Research on key technologies such as image processing and detection system classification and recognition has solved the problem of steel surface defect detection system. Defect segmentation takes a long time, there are many feature sizes, and the classification result is low.
Generally, we hope that the image captured by the camera is clear and noise-free [
Among them,
In multiplicative noise, the relationship is as follows:
Some infection spots with small gray values (close to black) or large gray values (close to white) appear in the image, and dark and bright spots similar to pepper and salt particles appear in the image, calculated as follows:
In the formula,
Formula (
In the formula,
In nanomaterials, when the particle size reaches a certain physical characteristic size, the energy levels of electrons adjacent to the metal Fermi plane change from an almost continuous state to a discrete state, while the nanoparticles have discontinuous, higher-occupancy molecules.
After filtering the image, we must also evaluate the image quality [
Assuming that the
Then, the two-dimensional entropy of zone
The discriminant function that defines entropy is
The so-called threshold segmentation technology is to use a threshold in the picture to divide the entire picture into two parts, a black part and a white part, with one part as the target object and the other part as the background object [
In the process of detecting data image defects, if you want to request the optimal threshold, you must cross all pixel values in the grayscale range and calculate the amount of change. When the amount of calculation is large, the output will be very low [
Compared with the Otsu threshold segmentation, it is not necessary to determine an optimal threshold for each picture, which has a certain degree of adaptability [
In the detection process of steel plate surface defects [
Integrate the various small areas so that the steel plate defect target and background can be separated as a whole [
In view of the small defect segmentation problem where the contrast between the defect target and the background is not obvious due to uneven illumination of the steel plate surface image and the reflection of the steel plate itself, the steel plate defect target segmentation algorithm proposed in this section fully considers the similarity of the defect target and the background pixel in the same area. Therefore, in the process of target segmentation, first, divide a small area; then, discharge whether there are defective targets in the small area; and finally, confirm whether the small defective target is a target [ The preprocessed steel plate surface image is divided into different small areas. In the small areas, each small area can be a background image or a combination of a background image and a defect target. It is better not to be all defective images Calculate the variance of each small area, and arrange them in an ordered sequence in the order of variance Set an initial threshold. If the variance change range is less than the initial threshold, the small area is determined to be a background image; if the variance change range is greater than the initial threshold, the small area is determined to have a defective target If it is determined that there is a defective target in a small area, the threshold within the variance is determined adaptively; if it is greater than the threshold within the variance, it is determined as the defective target; and if it is less than the threshold within the variance, it is determined as the background image After traversing each small area, each small area is divided into two parts: the defect target and the background. During the whole process, some backgrounds will be mistaken for the surface defect target of the steel plate, so it must be eliminated. The defect target has a certain path. Connected domains exist. This article takes the current pixel as the center and sets a
The steel plate surface defect detection system based on machine vision is mainly composed of optical lighting system, industrial camera, image acquisition system, image processing system, terminal computer, and data management system [
System flow chart.
The three doctors with rich work experience in this hospital are comprehensively judged for judging the degree of treatment. If there is a dispute, the result can be selected through discussion. We can regard the gray level of an image as a two-dimensional gray density function; then, a gray matrix can be used to describe the image moment features. We select a steel plate; take the invariant moment features of the cracks, holes, scratches, oil stains, and other images on it; extract the data results; and analyze them. Then, we read the texture features of these defect images again, extract the data results, and analyze them. Taking the image of steel plate without bonding defects and adding different concentrations of salt and pepper noise as the research object, 4 kinds of filtering methods are used to denoise mixed experiments.
There are many ways of data standard processing, but different data standardization methods will have a certain impact on the evaluation results of the system. For the positive index standardization method,
For the negative index standardization method,
After standardizing the data, using the principal component analysis of nonlinear logarithmic centering, the processing steps of logarithmic transformation and row vector centering are
SPSS23.0 software was used for data processing, and the count data was expressed as a percentage (%),
Reliability refers to the stability and reliability of the questionnaire [
Summary table of reliability test results.
Category | Index combination | Alpha coefficient ( |
---|---|---|
Crack | The crack itself | 0.8227 |
Crack right | ||
Crack left | ||
Crack transpose | ||
Hole | The hole itself | 0.8742 |
Hole moves right | ||
Hole moves left | ||
Hole transposition | ||
Bruise | Bruise itself | 0.7663 |
Bruise shifts right | ||
Bruise shifts left | ||
Bruise transposition | ||
Inclusion | Inclusion itself | 0.7414 |
Inclusion shifts right | ||
Inclusion shifts left | ||
Inclusion transposition |
It can be seen from Table
Feature extraction, as a key link in the image processing process in the system, is an important process to ensure the practicability of the system and the accuracy of defect recognition [
Invariant moment feature extraction data table.
Sample defect type | IM1 | IM2 | IM3 | IM4 | IM5 | IM6 | IM7 |
---|---|---|---|---|---|---|---|
The crack itself | 3.9843 | 17.7438 | 7.6313 | 2.6201 | 4.3140 | 5.4762 | 12.6321 |
Crack right | 3.3242 | 10.8950 | 7.3796 | 2.3861 | 4.6513 | 1.3793 | 7.2141 |
Crack left | 3.5124 | 11.3472 | 9.3724 | 2.3912 | 2.6724 | 1.6742 | 6.3241 |
Crack transpose | 3.9843 | 17.7438 | 9.1313 | 2.1201 | 3.8140 | 0.4762 | 12.6321 |
The hole itself | 5.7260 | 31.5000 | 27.8151 | 27.1595 | 15.6138 | 19.4741 | 9.6081 |
Hole moves right | 5.7260 | 31.5000 | 27.8151 | 27.1595 | 15.6138 | 19.4741 | 9.6081 |
Hole moves left | 5.7260 | 31.5000 | 27.8151 | 27.1595 | 15.6138 | 19.4741 | 9.6081 |
Hole transposition | 5.7260 | 31.5000 | 27.8151 | 27.1595 | 15.6138 | 0.8770 | 9.6081 |
Bruise itself | 2.7234 | 4.7278 | 7.5965 | 6.4868 | 8.2521 | 9.7417 | 9.6433 |
Bruise shifts right | 2.5343 | 2.9342 | 1.2417 | 7.5237 | 3.8921 | 5.8660 | 5.4276 |
Bruise shifts left | 2.5621 | 3.1137 | 1.2887 | 45181 | 2.7710 | 2.7770 | 0.3873 |
Bruise transposition | 2.7234 | 4.7278 | 0.5965 | 2.4868 | 0.2521 | 2.7417 | 0.6433 |
Moment invariant feature extraction data analysis diagram.
It can be seen from Figure
Feature extraction, as a key link in the image processing process in the system, is an important process to ensure the practicability of the system and the accuracy of defect recognition. We gather the moment invariant features extracted from different types of defects according to Table
Moment invariant feature maps extracted from different types of defects.
It can be seen from Figure
As a global feature, texture is a ubiquitous but difficult to describe possibility in images. The texture attribute refers to the law of change from pixel level to gray level in an image, which is an irregular but normal feature in the macroscopic view. We extracted the texture features of the cracks, holes, scratches, and inclusion defects in the samples, and the extraction results are shown in Table
Defect image texture feature extraction table.
Defect sample | Crack | Hole | Bruise | Oily | Inclusion |
---|---|---|---|---|---|
Energy | 0.9547 | 0.9491 | 0.8798 | 0.9736 | 0.8473 |
Mean gray value | 0.0278 | 0.0121 | 0.0372 | 0.0232 | 0.1422 |
Gray mean square error | 0.3312 | 0.0378 | 0.6208 | 0.3564 | 1.0381 |
Gradient mean | 0.1639 | 0.0761 | 0.3942 | 0.1536 | 0.2010 |
Gradient mean square error | 2.6790 | 0.9712 | 4.0097 | 2.7180 | 4.9014 |
Gray entropy | 0.0540 | 0.0166 | 0.0719 | 0.0794 | 0.0359 |
Gradient entropy | 0.0534 | 0.0154 | 0.0565 | 0.0670 | 0.1673 |
Mixed entropy | 0.0796 | 0.0165 | 0.1185 | 0.1786 | 0.1451 |
Texture feature extraction map of defect images.
It can be seen from Figure
Here, our method is to standardize the extracted feature data. We choose the linear scale transformation method, the range transformation method, and the standard sample transformation method. These three methods are different. The standard values of the data after linear transformation are all in (within the range of 0, 1); when the positive and negative indicators are equalized to positive indicators, the optimal value is 1, the worst value is 0, and the larger the value, the better. When the maximum value of a positive indicator is 0, this method cannot be used to standardize the indicator; when the value of a reverse indicator is 0, this method cannot be used to standardize the indicator. The results are shown in Table
Depression data analysis table.
Sample number | ||||||
---|---|---|---|---|---|---|
1 | -0.6666 | -0.6961 | 0.2436 | 1.9268 | -0.3848 | -0.5962 |
2 | -0.7104 | -0.7108 | 0.5538 | 1.7370 | -0.3880 | -0.6295 |
3 | -0.6335 | -0.6378 | 0.4459 | 1.8837 | -0.3642 | -0.6141 |
4 | -0.8493 | -0.8457 | 1.3670 | 0.9165 | 0.1147 | -0.7533 |
5 | -0.7154 | -0.7567 | 0.8430 | 1.6722 | -0.2849 | -0.6791 |
6 | -0.7378 | -0.7761 | 0.6645 | 1.7827 | -0.2274 | -0.6609 |
7 | -0.5272 | -0.5786 | 0.1737 | 1.9857 | -0.4264 | -0.5432 |
8 | -0.7264 | -0.7565 | 0.7528 | 1.6483 | -0.2781 | -0.6341 |
9 | -0.8715 | -0.8262 | 0.9458 | 1.4934 | -0.1535 | -0.6670 |
Standardized processing diagram of feature data.
It can be seen from Figure
Through four different strip steel surface defect detection system designs, work in the same environment at the same time to analyze the changes in detection accuracy. Data-type factors adopt independent sample
Analysis table based on the experimental results of image denoising.
Assignment variable | Proportion | |
---|---|---|
Mean filter | No salt and pepper added | 93.33% |
15% salt and pepper | ||
30% salt and pepper | ||
50% salt and pepper | ||
Gaussian filtering | No salt and pepper added | 94.44% |
15% salt and pepper | ||
30% salt and pepper | ||
50% salt and pepper | ||
Median filter | No salt and pepper added | 91.43% |
15% salt and pepper | ||
30% salt and pepper | ||
50% salt and pepper | ||
Median filtering based on partial differentiation | No salt and pepper added | 94.12% |
15% salt and pepper | ||
30% salt and pepper | ||
50% salt and pepper |
Analysis table based on the experimental results of image denoising.
Filtering algorithm | 15% salt and pepper noise | 30% salt and pepper noise | 50% salt and pepper noise | |||
---|---|---|---|---|---|---|
MSE | PSNR | MSE | PSNR | MSE | PSNR | |
Mean filter | 46.276 | 32. 2271 | 31. 2271 | 33.3695 | 85.972 | 28. 5943 |
Gaussian filtering | 42.873 | 36.1378 | 37.1738 | 37.2514 | 77.348 | 33. 1626 |
Median filter | 29.274 | 49.2472 | 50.2427 | 45.3223 | 39.734 | 41.5851 |
Median filtering based on partial differentiation | 18.396 | 56. 3471 | 57. 4317 | 45.2558 | 22.946 | 44.2432. |
Analyze the graph according to different algorithm experiments.
Figure
In order to compare different filtering effects more intuitively, this article uses signal-to-noise ratio and variance to evaluate the image quality after skipping. Comparing the maximum signal-to-noise ratio and image fluctuation after denaturation under different salt and pepper noise concentrations, it can be compared that the filtering effect studied in this paper is the best. Through experiments, the maximum signal-to-noise ratio and change value of each filter under different salt and pepper noise concentrations are obtained, as shown in Figure
Analyze the graph according to the signal-to-noise ratio and variance value.
It can be seen from Figure
An effective threshold segmentation algorithm for coefficient of variation is proposed. This algorithm overcomes the disadvantage of using only one threshold per iteration in repeated threshold segmentation. It also utilizes the sliding window used in adaptive threshold segmentation, which reduces the amount of calculation and improves detection efficiency. The system structure of the whole system is designed, and the system is divided into imaging module, fault detection software module, and storage management module. According to different principles of detection methods, different image capturing methods are given. The overall process of defect detection software is designed, and the design and implementation of basic software units are introduced in detail. This paper analyzes the design of the steel strip surface defect detection system based on machine vision. According to the application requirements of machine vision technology, based on the surface quality of strip steel, the detection of surface defects is adjusted and optimized for the design of this article. Experiments prove that the method designed in this paper is very effective. We hope that the research in this article can provide a theoretical basis for the design method of steel surface defect detection system based on machine vision.
At present, most retinal blood vessel extraction methods are mainly used for normal retinal images. Therefore, when applied to a large range of lesion images, it is difficult to accurately extract blood vessels due to the interference of lesions and other nonvascular structures, and a large number of nonvascular structures cannot be filtered. In addition, this article focuses on vascular bone extraction and vascular structure segmentation methods suitable for retinal imaging. Analyze the characteristics of defects, and find the entry point for detecting defects. According to the edge feature, uneven texture feature, and uneven defect feature, the defect detection method based on edge feature, the defect detection method based on unequal texture feature, and the defect detection method based on irregular feature are proposed. The techniques and theories on which this method relies are introduced and discussed.
The steel industry occupies an important position in China’s economic industry. It can be regarded as the main core of manufacturing. This is an intensive industry that has accumulated a lot of capital and energy. Although China produces a large amount of steel every year, the quality of various steel products in China is obviously not as good as that of developed countries. In the quality of steel products, the importance of surface quality is self-evident, but improving quality is always difficult. Improving the surface quality of steel is of great significance to improving market competitiveness. Accurate detection of steel plate surface defects and the establishment of a steel plate surface defect evaluation system are important conditions for improving the quality of steel plates. This article makes full use of the image data of steel surface defects and introduces artificial intelligence methods to classify and identify steel defects and target detection. It solves the problem of long iteration time of ant colony optimization algorithm and particle optimization algorithm. The algorithm is easy to fall into local optimization, improves the classification accuracy of the support machine, optimizes the optimization process, and is made of steel.
The data underlying the results presented in the study are available within the manuscript.
The authors declare that they have no conflicts of interest.