CN108846831B

CN108846831B - Band steel surface defect classification method based on combination of statistical characteristics and image characteristics

Info

Publication number: CN108846831B
Application number: CN201810524655.7A
Authority: CN
Inventors: 蔡炜; 叶理德; 欧燕; 梁小兵; 夏志
Original assignee: Wisdri Engineering and Research Incorporation Ltd
Current assignee: Wisdri Engineering and Research Incorporation Ltd
Priority date: 2018-05-28
Filing date: 2018-05-28
Publication date: 2021-09-28
Anticipated expiration: 2038-05-28
Also published as: CN108846831A

Abstract

The band steel surface defect classification method based on the combination of the statistical characteristics and the image characteristics comprises the following steps: collecting a set of labeled training sample sets; s2, extracting corner points of each defect sample image and description of each corner point; carrying out unsupervised learning clustering on the angular point descriptions of all the defect sample images in a K-dimensional space; combining the M-dimensional image feature vector and the N-dimensional statistical feature vector of the defect sample to form an M + N-dimensional feature vector of the defect sample; and (3) performing supervised learning training on the M + N-dimensional feature vector by using a self-adaptive lifting tree training method, training a self-learning classifier B, and outputting a classification result of the defect. The method distinguishes different defects, and improves the classification accuracy of the surface defects of the strip steel; when the surface defects are detected in real time on line, the self-learning classifier is used for automatically and accurately classifying the detected defects, and classification rules are obtained through supervised learning by using a machine learning technology without manually relying on manual input.

Description

Band steel surface defect classification method based on combination of statistical characteristics and image characteristics

Technical Field

The invention belongs to the field of strip steel surface defect detection systems in the metallurgical industry, particularly relates to the field of surface defect classification, and particularly relates to a strip steel surface defect classification method based on combination of statistical characteristics and image characteristics.

Background

The surface defect of the strip steel is an important factor influencing the surface quality of the cold-rolled strip steel, and directly influences the appearance and the service performance of a final product. The surface detection system scans the surface of the strip steel by using a camera sensor to obtain a two-dimensional image of the surface of the strip steel, and detects and classifies surface defects by using a machine vision technology, the accuracy rate of defect detection by various advanced algorithms is up to 98 percent at present, the production requirement is basically met, but the classification accuracy rate of the defects is always strong and is generally only about 80 percent, and the requirements of steel enterprises on product quality control and production process improvement cannot be met.

Disclosure of Invention

The technical problem to be solved by the invention is to provide a band steel surface defect classification method based on the combination of statistical characteristics and image characteristics aiming at the defects of the existing band steel surface defect classification, so that the classification accuracy of a band steel surface quality detection system on the band steel surface defects is improved; when the surface defects are detected in real time on line, the self-learning classifier is used for automatically and accurately classifying the detected defects, and classification rules do not need to be manually input.

The technical scheme adopted by the invention for solving the technical problems is as follows:

the band steel surface defect classification method based on the combination of the statistical characteristics and the image characteristics specifically comprises the following steps:

s1, collecting defect samples, classifying each defect sample according to the statistical characteristics of the defect samples, including area, perimeter, length-width ratio, position and appearance of the image of the defect sample, generating a group of training sample sets with marks, and training a self-learning classifier on the training sample sets, wherein the training sample sets are assumed to contain W defect samples and C-type defects;

s2, extracting corners of each defect sample image and each corner description, wherein the corner descriptions reflect the relationship between the corners and neighborhood pixel points, and the corner descriptions are assumed to be K-dimensional vectors;

s3, performing unsupervised learning clustering on the corner descriptions of all the defect sample images in a K-dimensional space, wherein the clustering number is M, the clusters are L (1), L (2) and L (3.. L (M)), and the category centers are LC (1), LC (2) and LC (3.. LC (M);

s4, calculating the probability distribution (image characteristics) of the corner point description of each defect sample in the M classes, namely counting the number of the corner point description of each defect sample belonging to the classes L (1), L (2) and L (3).. L (M), and arranging the corner point description into an M-dimensional vector, thereby forming the M-dimensional image characteristic vector of the defect sample;

s5, extracting N statistical characteristics (such as area, perimeter and position) of each defect sample, wherein the statistical characteristics form an N-dimensional statistical characteristic vector of one defect sample;

s6, combining the M-dimensional image feature vector and the N-dimensional statistical feature vector of the defect sample to form an M + N-dimensional feature vector of the defect sample;

s7, performing supervised learning training on the W M + N-dimensional feature vectors with the classification marks by using an adaptive lifting tree training method, training a self-learning classifier B, automatically classifying the detected defects (distinguishing different defects) by the self-learning classifier B, and outputting classification results of the defects.

According to the scheme, when the real-time online detection is carried out, the self-learning classifier B is utilized to classify the new defects by adopting the following steps when the new defects are detected, and the method mainly comprises the following steps:

i) extracting angular points of the defect image and K-dimensional description of each angular point, and calculating probability distribution of the K-dimensional description of the angular points in the categories of L (1), L (2) and L (3).. L (M) to serve as M-dimensional image feature vectors of the defect;

ii) calculating an N-dimensional statistical feature vector for the defect;

and iii) merging the M-dimensional image feature vector and the N-dimensional statistical feature vector of the defect to form an M + N-dimensional feature vector of the defect, and inputting the M + N-dimensional feature vector into a self-learning classifier B for prediction to obtain a classification result of the defect.

Compared with the prior art, the invention has the following beneficial effects:

1. combining the defect statistical characteristics with the image characteristics to form a feature vector of the defect, training a self-learning classifier in the feature vector space, distinguishing different defects, and improving the classification accuracy of the strip steel surface quality detection system on the strip steel surface defects;

2. when the surface defects are detected in real time on line, the self-learning classifier is used for automatically and accurately classifying the detected defects, and classification rules are obtained through supervised learning by using a machine learning technology without manually relying on manual input.

Drawings

FIG. 1 is a flow chart of a method for classifying defects on the surface of a strip steel based on the combination of statistical characteristics and image characteristics according to the present invention;

FIG. 2 is a diagram illustrating a classification process for detecting a new defect according to an embodiment of the present invention.

Detailed Description

The present invention will be described in further detail with reference to the accompanying drawings and examples.

The traditional classification of the surface defects of the strip steel is mainly realized by manually setting rules according to the statistical characteristics of the defects, for example, for punching defects, the rules such as circularity <1.1, mean value of gray values <40 and width <50mm can be generally set. However, it becomes difficult to manually establish rules for defects with more difficult description of specific characteristics, such as phosphorus spots, warping, etc. With the progress of machine learning technology, defect classification is carried out in a sample self-learning mode at the present stage, namely a group of samples which are marked with classification results manually are given, and then a computer is enabled to automatically find out the classification rules by using the machine learning technology. Generally, a defect is described by using statistical characteristics of the defect, and then a classifier is trained on the dimension of the statistical characteristics, where the statistical characteristics are some characteristics related to the position, shape, texture, gray level, topology, and the like of the defect, and mainly include: defect length, width, area, perimeter, aspect ratio, compactness, distance to strip steel boundary, center of gravity position, circularity, mean gray level, variance gray level, euler number, and the like. The statistical features are summaries and refinements of the defect information, which are to a large extent macroscopic manifestations of the defect features. The statistical characteristics are used, and when the macroscopic characteristics among defect categories are large in difference, the classification effect is good, such as distinguishing white spots from black spots, and distinguishing welding seams from holes. When the macroscopic features of the defects are very different, the classification accuracy is greatly reduced, such as distinguishing the defects of scabs and contusions. The corner features of the image better describe the detail and texture information of the image, in practical application, the corner features of the image are often used for face detection and pedestrian recognition, and the corner features are closer to the visual features of human beings, namely, the points with abrupt gray changes on the image are concerned more. However, if only the image corner feature is used for classifying defects, the defect classification is not enough, because the classification of many defects is not only related to the appearance of the defects, but also has a great relationship with the statistical characteristics of the defects, the statistical characteristics of the defects are combined with the corner features of the defect image, the union of the defect features and the corner features is taken as the characteristic vector of the defects, and a machine learning algorithm is used for automatically training a classifier to classify the defects.

Referring to fig. 1, the method for classifying the surface defects of the strip steel based on the combination of the statistical characteristics and the image characteristics, which is disclosed by the invention, combines the defect statistical information and the image information to train a self-learning classifier, and automatically classifies the detected defects by using the self-learning classifier, and specifically comprises the following steps:

s1, collecting a large number of typical defect samples, classifying each defect sample by a professional according to the statistical characteristics of the defect sample, such as area, perimeter, length-width ratio, position and the like, and the appearance of the image of the defect sample, training a self-learning classifier on the training sample set, and if W defect samples and C defects are contained, generating a group of marked training sample sets, wherein the defect classification can be carried out manually only by requiring the statistical characteristics of the image of the defect sample to correctly determine the category of the defect;

s2, extracting corners of each defect sample image and K-dimensional description of each corner, wherein FAST corners can be selected as a corner detection method in the step, each corner is an important local feature of the image and determines the shape of a target in the image, so that the method has important application in image matching, target description and identification, and the corners are positions with severe gray level change in a two-dimensional space of the image and are pixel points with obvious difference with surrounding adjacent points; because the camera is used for shooting at a fixed position in defect detection, the interference of background light is small, and the angular points are not required to have rotation and scaling invariance, the FAST angular point with the highest detection speed can be used;

s3, performing unsupervised learning clustering on the angular point descriptions of all the defect sample images in a K-dimensional space, wherein the clustering number is M, the clusters are respectively L (1), L (2) and L (3.. L (M), and the category centers are LC (1), LC (2) and LC (3.. LC (M);

s4, calculating and generating probability distribution of the corner point description of each defect sample in the M classes, namely counting the number of the corner point description of each defect sample belonging to the classes L (1), L (2) and L (3.. L (M)), thereby forming an M-dimensional image feature vector of one defect sample; because the numbers of the corner points of different samples are not consistent, the features must be compressed into a vector for description, the steps S3 and S4 cluster the description vectors of all the corner points in the dimension of the description vectors, and obtain M feature centers, thereby converting the corner point descriptions of a plurality of vectors of the image into a single vector description of probability distribution of the corner point descriptions on the M cluster centers, which can be regarded as a feature dimension reduction process;

s5, when each defect sample is detected, calculating N statistical characteristics of the defect, wherein the statistical characteristics form an N-dimensional statistical characteristic vector of the defect sample; the step is actually completed in real time during defect detection, because position characteristics and the like can be obtained only during defect detection, and position information cannot be obtained only from a defect sample image;

s6, combining the M-dimensional image feature vector and the N-dimensional statistical feature vector of the defect sample to form an M + N-dimensional feature vector of the defect sample; the statistical characteristics and the image characteristics of the defect sample are integrated, and the defect is represented by an M + N-dimensional characteristic vector containing the statistical characteristics and the image characteristic information;

s7, performing supervised learning training on the W M + N-dimensional feature vectors with classification marks by using a self-adaptive lifting tree training method, training a self-learning classifier B, automatically classifying the detected defects by the self-learning classifier B, distinguishing different defects and outputting classification results of the defects; a common classifier training method includes: random forests, self-adaptive lifting trees, support vector machines and the like; the features used for classification comprise statistical features and image features, the statistical features and the image features belong to different physical quantities, have large numerical difference and belong to mixed data, normalization processing of the data is not needed when random forests and self-adaptive tree promotion are used, classification can be carried out only after the data are normalized by using a support vector machine, and data preprocessing is complex; when the generalization capability of the classifier is actually evaluated by adopting a cross validation technology, the classification accuracy of the adaptive lifting tree on the test set is the highest and is about 85%, so that the classifier is trained by using an adaptive lifting tree training method.

Referring to fig. 2, in real-time online detection, when a new defect is detected, the defect needs to be described as an M + N-dimensional feature vector according to the above steps and classified by using a trained self-learning classifier B, which mainly includes the steps of:

i) extracting corners of the defect image and K-dimensional description of each corner, calculating probability distribution of the corner in the categories of L (1), L (2), L (3.. L (M)), calculating the distance from the feature description C of each corner to M cluster centers LC (1), LC (2), LC (3.. LC (M)), and then finding out the category corresponding to the center with the minimum distance, if C is closest to the center point of the category LC (x), the corner belongs to the categories of L (x), adding 1 to the value of an image feature vector I (x), and traversing the feature description of each corner of the defect image to obtain an M-dimensional image feature vector of the defect;

ii) calculating an N-dimensional statistical feature vector for the defect;

The above description is a preferred embodiment of the present invention, but the present invention should not be limited to the disclosure of the embodiment and the drawings. Therefore, it is intended that all equivalents and modifications which do not depart from the spirit of the invention disclosed herein are deemed to be within the scope of the invention.

Claims

1. The band steel surface defect classification method based on the combination of the statistical characteristics and the image characteristics is characterized by comprising the following steps:

s4, calculating the probability distribution of the corner point description of each defect sample in the M classes, namely counting the number of the corner point description of each defect sample belonging to the classes L (1), L (2) and L (3.. L (M)), and arranging the corner point description into an M-dimensional vector, thereby forming an M-dimensional image feature vector of the defect sample;

s5, extracting N statistical characteristics of each defect sample, wherein the statistical characteristics form an N-dimensional statistical characteristic vector of the defect sample;

s7, performing supervised learning training on the W M + N dimensional feature vectors with the classification marks by using a self-adaptive lifting tree training method, training a self-learning classifier B, automatically classifying the detected defects by the self-learning classifier B, and outputting the classification results of the defects.

2. The method for classifying the surface defects of the strip steel based on the combination of the statistical characteristics and the image characteristics as claimed in claim 1, wherein when a new defect is detected in real time on-line detection, the self-learning classifier B is used for classifying the new defect by adopting the following steps, which mainly comprise the following steps:

ii) calculating an N-dimensional statistical feature vector for the defect;