Defect feature recognition method of glass fibre-reinforced structure based on visual image analysis

HUANG Jingde

(Zhuhai College of Science and Technology, Zhuhai 519041， China)

Abstract： Glass fibre-reinforced (GFR) structure is extensively used in radome, spoiler and some other equipment. In engineering practice, due to the influence of wear, aging, impact, chemical corrosion of surface structure and other factors, the internal structure of this kind of structure gradually evolves into a defect state and expands to form defects such as bubbles, scratches, shorts, cracks, cavitation erosion, stains and other defects. These defects have posed a serious threat to the quality and performance of GFR structure. From the propagation process of GFR structure defects, its duration is random and may be very short. Therefore, designing a scientific micro defect intelligent detection system for GFR structure to enhance the maintainability of GFR structure will not only help to reduce emergencies, but also have positive theoretical significance and application value to ensure safe production and operation. Firstly, the defect detection mechanism of GFR structure is discussed, and the defect detection principle and defect area identification method are analyzed. Secondly, the processing process of defect edge signal is discussed, a classifier based on MLP is established, and the algorithm of the classifier is designed. Finally, the effectiveness of this method is proved by real-time monitoring and defect diagnosis of a typical GFR structure. The experimental results show that this method improves the efficiency of defect detection and has high defect feature recognition accuracy, which provides a new idea for the on-line detection of GFR structure defects.

Key words： glass fibre-reinforced (GFR) structure； multi-layer perceptron (MLP)； machine vision； defect detection

0 Introduction

Glass fibre-reinforced (GFR) structure with high specific strength and good impact resistance is widely used in rudder, radome, spoiler and torpedo tube of weapon system. The structure of this kind of equipment has tiny size and smooth surface. However, its surface quality is so brittle that the quality of surface is unstable. Hence, the micro defects such as bubbles, scratches, shorts easily appear. When the equipment is under cyclic load, it is very easy to cause or induce functional failures[1]. At present, the percussion testing method, non-destructive testing method and ultrasonic testing method for this type of equipment cannot effectively analyze failure propagation mechanism and predict failure time[2-5]. From the development process of surface defects of GFR structure, its duration is random and might be very short, resulting in difficult detection and diagnosis. Therefore, it is significant to realize real-time and accurate detection of the above surface defects, timely monitor the abnormal situation of GFR structure, discover the abnormal situation and timely maintain. In recent years, the appearance of defect detection technology based on machine vision has greatly improved the efficiency of production and operation. This technology reduces the influence of operation conditions and subjective judgment on the accuracy of detection results, so that structural defect detection can be realized better and more accurate. Besides, product defects can be identified rapidly[6-9]. At present, although GFR structure defect detection method has achieved a lot of results, most studies focus on the causes of defects, detection and processing algorithms. The module design is relatively single, and the overall classification algorithm is complex and the recognition rate is low. With the gradual improvement of technology, the research focus is gradually inclined to the field of surface engineering detection, focusing on solving the following problems. First, how to use the image acquisition system to accurately collect the structural features of the object surface； Second, how to accurately identify and segment the defect area； Third, how to enlarge the defect characteristics and realize the classification simply and correctly.

1 Defect detection mechanism

1.1 Defect detection principle

The main defects of GFR structure are scratches, bubbles and shorts. Because of its compact material and fine surface structure, the gray value of the non-defective GFR structure image is basically the same, and there will be no light path deflection when the light enters. If there are defects in its interior, the light path will deflect at the edge of the bubbles, which results in obvious deviation of the gray value； If the scratch is caused by external force, since the surface roughness is not consistent, the light refraction will be weakened, and the gray value of the defect will be lower； At the same time, in the process of production and transportation of GFR structure, due to the difference of process level, the surface may be uneven, and the light refractive index may also be affected. Therefore, if there are defects in the GFR structure, the refractive index of light can be used to judge whether there is obvious change in the difference of plane gray level, which can reflect the problem of glass defects as a whole[7-10]. According to the principle of charge-coupled device (CCD) detection, the surface of non-defective GFR structure is smooth, even, and the gray value distribution is relatively average. When the incident light enters, the light path is shown in Fig.1.

(a) Non-destructive glass structure

(b) Containing impurities such as stains

(c) Containing impurities such as bubbles

It can be seen from Fig.1 that the glass without defects will not deflect when the light is emitted, and the CCD can recognize the light with uniform refraction； If there are impurities such as stains in the glass, the incident light will be absorbed, and the refraction will be weakened. Because the light received by the CCD target will be reduced, the gray value of the image will be reduced too. If there are defects such as bubbles or scratches in the structure, although there will be light refraction in the defects, some of the light will be directly transmitted. Therefore, the gray value of the edge of such defects will be significantly lower than the gray value of the interior. Besides, the size of bubbles and other defects will also affect the light transmittance. The physical expression is the indefinite change of gray value.

1.2 Defect region recognition method

Huinvariant moment is a classical method in region recognition. For two-dimensional imagef(x,y), the standard momentmpqis defined as

(1)

wherep,q=0,1,2,…,n.

Using the moment of the target, its center of gravity coordinates can be calculated by

(2)

Then the central momentμpqof imagef(x,y) is defined as

(3)

For digital images, the moment can be expressed as

(4)

The central moment can be expressed as

(5)

In the image,i,jcorresponds tox,ycoordinates, respectively.

Iff(i,j) is a binary image, the moment can be calculated by

(6)

The advantage ofHuinvariant moment is that it requires less computation. Moment method is a concise mathematical representation, which has been widely used in target region recognition.

2 Defect edge signal processing method

2.1 Gaussian filtering model

In the engineering practice, the collectedGFR structure image will certainly produce noise interference. Before the image processing, it is generally necessary to use filtering mechanism to improve the imaging quality of the original image in order to improve the correlation of strong defect features. The noise produced in the process of GFR structure image acquisition is mainly white noise conforming to Gaussian distribution function. Therefore, the Gaussian filtering model has a strong processing effect on this kind of noise. The gray value of each pixel is replaced by the weighted average of the gray values of itself and surrounding pixels. The formula is

(7)

whereG(x,y) is the image weight matrix andσis its component on thex-axis andy-axis.

2.2 Edge gradient determination method of defect images

Sobel operator is an operator that uses local difference to find edges[11]. It is mainly used to process the area of the image with 3×3 pixels. The two convolution kernels of Sobel operator are shown in Fig.2.

Fig.2 Convolution kernel templates of Sobel operator

In order to enhance the sensitivity of weak edges and improve the accuracy of edge detection, we choose Sobel operator to perform convolution operation through the original image and two groups of 3×3 template matrices to realize the approximate processing of horizontal pixels and vertical pixels. The horizontal gradientGxand vertical gradientGyare

(8)

(9)

whereAis the original image.

During convolution operation, we turn the template 180° first, and then calculate the gray value of the point by

(10)

TakingIas the marking matrix of adjacent pixels,

(11)

thenGxcan be expressed as

Gx=(a2+2a3+a4)-(a0+2a7+a6),

(12)

andGycan be expressed as

Gy=(a0+2a1+a2)-(a6+2a5+a4).

(13)

Finally, we get the gradient directionθis

(14)

2.3 Segmentation method of defect image

Considering that the background after GFR structure imaging is relatively single and the gray distribution of the defect area is relatively uniform, in order to control the error change, we use the maximum interclass variance method to segment the image[12]. Assuming that the background domain isC1and the target domain isC2, the greater the difference betweenC1andC2, the higher the accuracy of segmentation. For imagef(x,y), assuming that the optimal threshold isT, and the input image hasM×Npixels, then we can get the following results.

The proportion of backgroundN1in all pixels is

(15)

The proportion of backgroundN2in all pixels is

w2=1-w1.

(16)

The average gray valueω1of backgroundN1is

(17)

(18)

wherep(i) is the probability density of gray leveliin the image.

The average gray valueω2of the targetN2is

(19)

The average gray level of the image is

ω=w1ω1+w2ω2.

(20)

Therefore, the maximum interclass variance is

g=w1w2(ω1-ω2)2.

(21)

According to Eq.(21), if the optimal threshold is to be established, the variance between classes must be maximized, that is, the gray mean value of the two parts should have a large difference.

3 Process of defect feature recognition

3.1 Classifier model based on MLP

On account of the image difference information which is given by the image difference method is a single pixel, in engineering practice, it is difficult to achieve ideal results by employing difference images[13-14]. Therefore, it is necessary to design a scientific defect feature classifier to implement image segmentation, which includes two stages of classifier: learning and classification. The inputs of classifier are difference information which should possess a certain practical significance and an actual region. Considering that the neural network directly determines the hyperplane of segmentation between classes, in order to enable the classifier to distinguish classes that are not linearly separable, the processing units of the input layer, hidden layer and output layer of multi-layer perception (MLP) structure adopt the linear combination of calculating eigenvectors or the results of the previous layer, as shown in Fig.3.

Fig.3 An MLP processing unit

Firstly, each processing unit calculates the excitation value of linear combination of input values by

(22)

Secondly, the results are introduced into the nonlinear excitation function, and then we have

(23)

In this work, hyperbolic tangent function is used as the excitation function, and thus we have

(24)

The corresponding function image is shown in Fig.4.

Fig.4 Tanh function image

Considering that MLP can approximate any nonlinear continuous function with any accuracy, it can well solve the identification problem of nonlinear system. At the same time, MLP also has good self-learning ability. It can extract regular knowledge from input and output data during training and memorize it in network weights. Therefore, we can get

(25)

3.2 Classifier training process

In order to obtain the segmentation hyperplane classified by MLP, the weights of neural network should be adjusted by training[15]. Compared with the single-layer perceptron, the MLP neural network not only adds the hidden layer, but also adds the activation function after the weight summation. Therefore, the nonlinear classification ability of the MLP neural network is further improved. The structural parameters of MLP are shown in Table 1.

Table 1 Parameters of MLP classifier segmentation algorithm

The implementation process of segmentation algorithm is as follows:

1) The defect area is determined by image segmentation;

2) Create MLP classifier model: create_ class_ mlp()；

3) Add training samples to MLP classifier: add_ samples_ image_ class_ mlp ()；

4) Train MLP classifier: Train_ class_ mlp ()；

5) MLP segmentation of image: classify_ image_ class_ mlp ().

4 Example analysis

4.1 Design of experimental system

The detection system employs a dual channel acquisition scheme, which can obtain two channels of image data at the same time. The experimental object of the system is about 2 mm glass structure. The algorithm implementation relies on the joint development of Halcon visual library and Visual Studio 2015, and the UI is written with C#. The detection system consists of three modules: auxiliary control mechanism, signal acquisition system and image processing system. The auxiliary control mechanism mainly controls the movement of the detected object through the manipulator. The signal acquisition system is mainly composed of two channels. The switching between channel 1 and channel 2 is completed by the auxiliary control mechanism switching the high-speed camera. In channel 1, defects such as stains, scratches and bubbles are mainly collected. The lighting mode of stains and bubbles is blue backlight. Scratches are mainly low angle blue light. In channel 2, uneven defects such as shorts is mainly collected. The grating stripe screen is used for backlighting. The collected optical signal is converted to electrical signal by image processing system. Then electrical signal is converted to digital signal by analog to digital converter (ADC) and the digital signal is transmitted to image feature processing module. The feature processing module can extract strong feature parameter, and then classify different defects by algorithm. The results are displayed on human-computer interactive device. The whole process is completed at one cycle. The structure of the experimental platform is shown in Fig.5.

4.2 Analysis of results

In order to verify the real detection effect of defects, the defect area is calibrated, and the results of difference are shown in Fig.6. Figs.6(a) and (b) show the filtered gray distribution information and its change trend. Since the compactness of gray distribution can approximate the proportion of noise in the image, the gray distribution information after difference reflects good convergence, which is conducive to defect feature extraction and separation.

(a) Gray distribution before difference

(b) Gray distribution after difference

Experiments were carried out on the glass with defects, The scratches of transparent components are generally narrow in width and long in length. When the short aspect ratio is less than 0.5, the gray image generally presents texture features. If the short aspect ratio is greater than 0.5, the gray image presents quasi circle features. The sample processing results are shown in Fig.7.

Fig.7 Defect feature distribution

The classification results of defect features are shown in Table 2.

Table 2 classification results of defect features

The experimental results show that all the defects are classified correctly, and the small points can be identified accurately.The classification effect is excellent.

5 Conclusions

In this work, we analyze the texture image of image surface by using image difference method. The image of defect information is segmented by this method. Hence, transparent structure defect images can be classified according to their unique regional features. The results of defect classification provide a theoretical basis for the defect diagnosis of GFR structure, and thus improves the accuracy of defect feature recognition. It provides a new solution and technical support for the scientific diagnosis and prediction of latent and intermittent faults.

1) The defect detection mechanism of GFR structure is studied, and the differential treatment process of defect area is analyzed.

2) A classifier based on MLP is designed to improve the classification accuracy of defect features and avoid the problem of mis-judgment of defects.

3) The multi-channel image acquisition mode is adopted to reduce the omission of strong feature parameters and simplify the classification of defects.