Adaptive defect detection of 3D image of nano CT based on wavelet decomposition

Abstract

The defect detection of 3D image of nano CT under different interference has the phenomenon of prominent dislocation. Therefore, an adaptive detection method of 3D image defect of nano CT based on wavelet decomposition is proposed. Analyze the noise of three-dimensional image of nano CT, determine the mixed filtering of image sequence according to the different noise properties, evaluate the mixed filtering of image sequence, complete the preprocessing of three-dimensional image of nano CT, fuse the three-dimensional image of nano CT decomposed by wavelet after preprocessing, enhance the image after decomposition, and realize the defect adaptive detection through the characteristics of wavelet decomposition. The experimental results show that the design method can effectively detect the interference and solve the problems of traditional methods.

Keywords

Wavelet biotechnology computational materials science 3D image image defect adaptive detection

1. Introduction

In recent years, with the rapid development of microelectronics technology, the adaptive detection performance of image defects has been greatly improved. The scanning mode has developed from two-dimensional fault scanning to three-dimensional cone beam scanning. It has the ability to scan hundreds of faults at the same time and uniform spatial resolution. New image principles and high-resolution technologies have also made breakthroughs, providing us with the ability to process three-dimensional images [1]. However, due to the limitation of image processing methods and the influence of CT three-dimensional scanning mode, the existing detection technology research is mainly focused on the three-dimensional analysis of nano CT three-dimensional image. The purpose of 3D image analysis of nano CT is to detect 3D defect information, extract defect parameters and identify defect types. If the 3D image model is constructed by the 3D image sequence of nano CT with isotropic characteristics, then the image sequence analysis technology can be used for adaptive detection and analysis [2].

Compared with ordinary images, medical images are fuzzy and uneven in nature. First of all, medical images are ambiguous in gray scale, and the signal value will change greatly in the same tissue. Secondly, due to the local volume effect, voxels on a boundary often contain two substances at the same time. In addition, due to technical reasons, the noise often blurs the high-frequency signal of the object edge. The wavelet decomposition method can recognize the mapping relationship between 3D image features and 3D shape features of nano CT, and realize 3D defect detection. In this paper, an adaptive defect detection method based on wavelet decomposition is proposed. Wavelet decomposition is used to fuse the preprocessed 3D images of nano CT to realize adaptive defect detection. The experimental results show that the edge detection of medical image based on this method has better noise suppression ability and complete edge preserving characteristics.

2. Image preprocessing

The existence of noise not only reduces the image quality, but also hinders the acquisition of useful information. The three-dimensional original slice image of nano CT has the characteristics of low gray contrast, fuzzy edge, and more noise information. Sometimes, the defect features are submerged by too much noise information, which affects the effect of analysis and evaluation of detection samples based on slice image [3]. In order to extract the defect features correctly, it is very necessary to preprocess the slice image. The main purpose of preprocessing is to remove the noise information, protect the image details, reduce the misjudgment caused by noise, and improve the reliability of defect detection. In this chapter, by analyzing the nature of 3D image noise of nano CT, combined with the existing filtering algorithm and the characteristics of image sequence, the image noise of 3D image sequence of nano CT is removed.

2.1 Noise analysis of 3D image of nano CT

The process of 3D image imaging of nano CT is to excite X-ray through high voltage electric signal, get X-ray signal through the object on the receiving screen, convert the ray signal into visible light in the receiving screen, and get gray-scale projection image through the intensity of light [4]; The projection image is then filtered and reflected, and the 3D reconstruction algorithm gets the slice image of the object. After the conversion of a series of photoelectric signals, there is inevitably noise in the final CT slice image.

It can be seen from the principle of 3D image imaging of nano CT that many factors will affect the performance of CT image, including: X-ray spectrum, photon scattering, focus blur, scanning geometry path, number of projection images, probe misalignment, detector bad spots, detector pixel size, reconstruction algorithm, quantum noise and electronic noise, etc. In particular, the combination of geometric errors and human factors makes all kinds of noise appear at the same time, which seriously affects the quality of 3D image of nano CT.

There are three main sources of noise in 3D image of nano CT: quantum noise, fuzzy noise of flat panel detector and additive system noise [5].

Under the influence of X-ray spectrum, photon scattering and noise source, there are prominent dislocations in the defect detection of three-dimensional CT images under these different interferences.

2.2 Image sequence hybrid filtering

The noise information in the image has a great influence on the subsequent defect detection. In this chapter, the nature of noise in the 3D image of nano CT is analyzed, and then combined with the existing classical filtering algorithm, a filtering method suitable for the 3D image sequence of nano CT is given.

In this paper, the simplest nonlinear filtering method is used. Median filter is widely used in image denoising because it can smooth the noise and protect some details in the image [6]. The basic principle of median filtering is to replace the value of a point in a digital image or digital sequence with the median of all points in a neighborhood of that point.

Let $\{{x({i,j}),({i,j})\in I\times I}\}$ represent the gray value of each point in the digital image, $r$ indicates that the sliding window needs to be designed, traversed and assigned successively. And the three-dimensional median filtering with $s$ as the filtering window is defined as:

$\displaystyle y({i,j})=\textit{median}\{{x,({i,j})}\}=\textit{median}\{{x,({i+% r,j+s})\in s}\}$ (1)

Figure 1.

Schematic diagram of the four filtering directions.

The window of three-dimensional median filter can be square, circular or cross shaped. The performance of standard median filter to remove impulse noise is greatly affected by the size of filter window. There is a certain contradiction between suppressing image noise and protecting details. The filter window is small, which can better protect some details in the image, but the ability of filtering noise will be limited [7]; The filter window is large, which can strengthen the ability of noise suppression, but the ability of protecting details will be weakened, sometimes it will filter out some important details in the image, such as thin lines, sharp edges and corners, so as to destroy the geometric structure of the image. Although the classical median filter is very effective in smoothing the impact noise and keeping the image details well, it will destroy many good image details regardless of whether the pixels are good or not. In order to solve this problem, many researchers put forward many adaptive algorithms [8, 9], this paper uses adaptive median filtering algorithm. The basic idea is to consider the directionality of the image, use four directions to calculate separately, use the difference between the four sub images and the original image to generate the weight, and finally use the four sub images to weight the final filtered image. Taking 5 $\times$ 5 window as an example, the original image is respectively median filtered in four directions, as shown in Fig. 1. The values of the four sub images are obtained as follows.

2.3 Filtering results and evaluation

Filter and preprocess the Fig. 1. The window size is 3 $\times$ 3 and 5 $\times$ 5 corresponding parameters NMSE, PSNR and filtering time t are shown in Table 1. The different angles in Fig. 1 are respectively the results obtained by using 5 $\times$ 5 window template instead of linear filtering, nonlinear filtering and hybrid filtering algorithm.

Table 1
Filtering performance parameters

Template size	The performance parameters	The linear filter	Nonlinear filtering	Hybrid filter
3 $\times$ 3	NMSE	0.00259	0.0080	0.00129
	PSNR	30.0544	35.2044	33.069
	T (S)	1	1	3
5 $\times$ 5	NMSE	0.00925	0.00391	0.00792
	PSNR	24.324	28.2625	25.2523
	T (S)	6	1	9

In the same template, the linear filtering method has the largest NMSE value and the smallest PSNR value [10], that is to say, it has strong denoising ability to select this value, but poor ability to protect details; While the nonlinear filte9ring has poor denoising ability and strong ability to protect details. In this paper, the method of transferring noise variance between layers is used to make noise estimation more accurate, improve the accuracy of filtering and the ability of image detail protection. Combining the advantages of linear filtering and nonlinear filtering, the ability of removing noise and protecting details are balanced.

3. 3D image fusion of nano CT based on wavelet decomposition

3.1 3D image fusion rules of nano CT based on wavelet decomposition

The theoretical basis of fusion rules is that after wavelet decomposition, the low-frequency band represents the approximate part of the image, and the high-frequency band represents the details of the image. However, the coefficient of high frequency band fluctuates around the zero value. The larger the absolute value is, the more intense the brightness change is. That is to say, it may contain important image information of three-dimensional image of nano CT, such as the edge of three-dimensional image of nano CT, linear features in three-dimensional image of nano CT and the edge of region. In addition, there are significant differences in the high-frequency details of the three-dimensional images of the same scene obtained by different sensors, while the coefficient values of the low-frequency approximate parts are not significantly different.

According to the characteristics of 3D image of nano CT after wavelet decomposition, there are a lot of research on fusion rules in the previous literature. According to the characteristics of each fusion rule, it can be roughly divided into three categories: one is the rules based on the global characteristics of the image [11]; The other is the rules based on the single pixel; The third is the rules based on the neighborhood characteristics of the pixel.

This kind of algorithm mainly includes: Based on the selection rules of median, absolute maximum, absolute minimum, simple weighted average, noise method, etc. To decompose the corresponding coefficients of two images in the fusion image, take the median value, the maximum value of absolute value and the minimum value of absolute value to the corresponding position of the fusion image; For simple weighted average, directly take a certain weight value to select the low-frequency part or high-frequency part of the image to be fused to the corresponding position of the fusion image; The noise extraction method is based on the high-frequency wavelet coefficient position of the noise in the wavelet decomposition. The size of the image is eliminated by a certain calculation judgment, and then fused into the corresponding image. The following describes the weighting method, selection method and noise extraction method.

The second kind is not suitable for the case where the source image can not be accurately registered and the spectral characteristics of the original image are quite different (such as SAR and visible image, infrared and visible image) because the single pixel rule does not consider the neighborhood information. The rule based on neighborhood feature is to take the energy, variance, gradient, distance and other features of neighborhood at a certain location, that is, the feature can reflect whether the location is an important information of the image, as a measure to guide the coefficient selection at the location, the neighborhood size is generally taken as 3 $\times$ 3, 5 $\times$ 5, 7 $\times$ 7 windows.

The third kind is based on the global feature, which mainly considers the global feature of the image. It is applied in the case that the spectral feature of the original image is quite different. The rules of global features take the variance, energy, gradient and distance of the whole image as a measure to guide the coefficient selection of the two images.

3.2 Wavebase selection of wavelet decomposition

In order to get better fusion effect of remote sensing image, wavelet selection is also very important. Different wavelet bases will produce different fusion results. Therefore, it is necessary to compare the commonly used wavelet bases to determine the best wavelet base in image fusion. At present, there is no good theoretical guidance for the selection of wavelet bases. Generally, the best wavelet bases are selected through a large number of experiments. The most commonly used wavelet transforms are Daubechies wavelet [12], biorthogonal wavelet, cubic B-spline wavelet. In this paper, we use these different types of wavelet to carry out experiments on remote sensing image fusion, and compare their effects. At the same time, referring to wavelet decomposition and lifting image fusion, we can find the best wavelet base for image fusion based on wavelet curve decomposition. At last, we choose the most commonly used db2, db4, sym2, sym4, bior3.3/rbio3.3 as wavelet basis to decompose and reconstruct the image [13]. Daubechies wavelet is generally abbreviated as dbN, N is the order of wavelet, the support area of wavelet function y(t) and scale function r(t) is 2N-1, and the vanishing moment of y(t) is n. Except for N $=$ 1, dbN has no symmetry (i.e. nonlinear phase). The wavelet function y(t) is generally determined by the scale function r(t), and the scale function is a trigonometric function. symN is the abbreviation of symlet. It is an approximate symmetric wavelet proposed by dbN and an improvement of db function. Biorthogonal wavelet is abbreviated as biorNr.Nd, where r represents reconstruction and d represents decomposition. The main feature of Biorthogonal wavelet is that it has linear phase, and its dual two wavelets are used for image decomposition and reconstruction respectively. The two-dimensional scaling function and wavelet function of sym4 are shown in Fig. 2.

Figure 2.

Analytical diagram of two-dimensional scale function and wavelet function.

3.3 3D image fusion of nano CT

Finally, there will be (3N $+$ 1) different subgraphs, including 3N high frequency subgraphs and one low frequency subgraph. The more layers of wavelet decomposition, the smaller the size of the corresponding layer image, so the image wavelet decomposition has a pyramid structure. N-level wavelet decomposition of two-dimensional remote sensing image is also a multi-resolution and multi-scale decomposition with direction. The image fusion method of wavelet decomposition is as follows:

If two original 3D nano CT images need to be fused by wavelet decomposition, the basic steps of fusion processing are as follows:

Wavelet decomposition of image: each original three-dimensional image of nano CT is transformed from spatial domain to frequency domain by wavelet transform. Each layer of wavelet decomposition includes 3N high-frequency subgraphs and one low-frequency subgraph:

Fusion processing: The corresponding coefficients of each decomposition layer are fused separately. Different fusion operators can be used to fuse different frequency components on each decomposition layer, and finally the fused wavelet coefficients are obtained; Image wavelet reconstruction: The wavelet coefficients obtained after fusion are inverse transformed by wavelet, that is, image reconstruction, and the reconstructed image is the fusion image.

Wavelet transform decomposition composite image is to use discrete wavelet transform to decompose each image of N images to be fused into M sub images, and then fuse m sub images from n images to be fused at each level to get the fused image of this level, as shown in Fig. 3.

Figure 3.

Fusion image process.

After wavelet decomposition, the frequency characteristics of the image are effectively separated. At this time, the high-frequency detail information of SAR image and the low-frequency spectral information of TM image can be fully used for fusion [14]. New wavelet coefficients can be generated through certain fusion rules at different levels, and then the final fusion results can be obtained through wavelet inverse transformation. In this way, the fusion image not only has the high-resolution characteristics of SAR image, but also retains the spectral characteristics of multispectral image.

4. 3D image enhancement based on wavelet decomposition

After the 3D image fusion of nano CT, it is necessary to enhance the features of 3D image of nano CT. Wavelet analysis can adaptively change the spatial and frequency windows of image signal, and also can carry out spatial-frequency localization analysis. Compared with Fourier analysis, it has better space frequency characteristics [15]. In the field of signal processing, wavelet analysis is often used to solve the practical problems, which mainly uses the characteristics of wavelet multi-resolution and multi-scale analysis. For image signal, the multi-resolution and multi-scale analysis can be achieved by changing the window size of processing signal in time-space domain and frequency domain.

When the 3D image of nano CT is transformed by wavelet, the wavelet forward transform corresponds to the decomposition process of the image, and the wavelet inverse transform corresponds to the reconstruction process of the image. After the image is decomposed by wavelet, the image is decomposed into different size, different direction and different frequency components. These components exist in the form of wavelet coefficients after the image is decomposed [16]. The special processing of the image can be realized by transforming the wavelet coefficients pertinently. It can be seen that wavelet analysis is to decompose the signal characteristics in the spatial or frequency domain and then analyze them.

In order to deal with the high-frequency and low-frequency components of the image signal in fractional order, a large band of the signal is gradually decomposed into a small band by virtue of the characteristics of wavelet multi-scale and multi-level decomposition. In the actual processing, in order to reduce the phase delay of the processed image and minimize the edge distortion of the processed image, the lifting format wavelet is chosen as the research object. It is a space-frequency analysis method [17], with fast operation speed, easy to realize forward, reverse and non-linear wavelet transform. At the same time, according to the classical Mallat tower wavelet decomposition theory, the image is divided into 3L $+$ 1 sub components, i.e. one low-frequency component and 3L high-frequency component after the three-dimensional discrete wavelet transform with L decomposition layers. The typical 3D wavelet decomposition process is shown in Fig. 4.

Figure 4.

Schematic diagram of 3D wavelet decomposition process.

It can be seen from Fig. 4 that the sub image signals of the original image after wavelet 1 layer decomposition are LL1, LH1, HLL and HHL respectively. Then LLl concentrates the main low frequency part of the original image and reflects the approximate information of the image [18]; L corresponds to the high frequency information in the horizontal direction; HL corresponds to the high frequency information in the vertical direction; HH corresponds to the high frequency information in the diagonal direction, which reflects the image details in 3 directions. The LL1 sub image signal obtained from the decomposition is also decomposed to obtain the sub image signals on the second layer.

Because the pixels in the image neighborhood have great correlation, the pixels in the sub image neighborhood obtained by wavelet decomposition also have great correlation, and have different characteristics of correlation. For the low-frequency subband image, the pixels in the contour neighborhood show great correlation; For the high-frequency subband image, the pixels have great directional correlation. If the special correlation of the sub image pixels obtained from wavelet decomposition can be used in image enhancement, more information can be enhanced in the image.

Let $\psi(t)$ be a wavelet function, and the compactness of $\psi(t)\in L^{2}(R)$ wavelet function has the following formula:

$\displaystyle C=\int{\frac{\psi(t)}{t}}dw\leqslant\infty$ (2)

The volatility of wavelet function makes the following formula hold:

$\displaystyle\int{\psi(t)dt=0}$ (3)

If scale and translation transformations are performed on $\psi(t)$ , they are:

$\displaystyle\psi(t)_{a,b}=\frac{1}{\sqrt{a}}\psi\left({\frac{t-b}{a}}\right)$ (4)

$a$ and $b$ are scale and translation factors respectively, and $\psi(t)_{a,b}$ is called wavelet basis function [19]. The essence of wavelet transform is realized by a set of wavelet basis functions and signal multiplication integral, including positive transform and inverse transform. $w_{f}$ : The name of the wavelet filter to use in the decomposition.

$C$ is the column vector. For one-dimensional continuous signal $f(t)\in L^{2}(R)$ , its positive transformation is as follows:

$\displaystyle w_{f}(a,b)=\frac{1}{\sqrt{a}}\int{\psi\left({\frac{t-b}{a}}% \right)}dt=\langle{f(t),\psi_{a,b}(t)}\rangle$ (5)

The inverse transformation expression is:

$\displaystyle f(t)=_{\psi}\int{\int{\psi_{a,b}}}(t)\frac{\textit{dadb}}{a^{2}}% =\langle{C_{\psi}^{-1},\psi_{a,b}(t)}\rangle$ (6)

After the inverse transform expression of the three-dimensional image of nano CT, the wavelet forward transform corresponds to the decomposition process of the image, and the wavelet inverse transform corresponds to the reconstruction process of the image. After the image is decomposed by wavelet, the image is decomposed into different size, different direction and different frequency components. These components exist in the form of wavelet coefficients after the image is decomposed. The special processing of the image can be realized by transforming the wavelet coefficients pertinently. It can be seen that in the process of wavelet decomposition, the wavelet frequency in the effective amplitude range of wavelet is a function, as shown in Fig. 5.

Figure 5.

Relationship of effective amplitude range functions of wavelet.

Take the first three terms of Eqs (2) to (5) to construct a type 3 $\times$ 3 fractional differential mask operator, as shown in Fig. 6. The 8 sub masks can respectively perform convolution filtering operations on the x-positive direction, x-negative direction, Y-positive direction, Y-negative direction, left upper diagonal direction, left lower diagonal direction, right upper diagonal direction and right lower diagonal direction of the $({x,y})$ force of the image $s({x,y})$ pixel, and improve the anti-rotation performance and isotropic effect of the operation [20]. For the 3D image of nano CT, the filtering operation is to move the template point by point for each pixel in the image to be processed, but at this time, the number of sub masks for each pixel is 8.

If the 3 $\times$ 3 asymmetric fractional order differential shown in Fig. 6 is applied to each pixel $({x,y})$ of the original image, the pixel with zero weight is ignored.

Figure 6.

Schematic diagram of 3 $\times$ 3 type asymmetric fractional order differentiation.

In order to make full use of the correlation between the image pixel and the gray value of the pixel in its surrounding area, the pixel point with constant coefficient “1” is divided into two parts: the pixel point with a distance of one pixel $\frac{-v}{5}$ and the pixel point with a distance of two pixels $\frac{({v^{2}-v})}{6}$ . In this way, the sum of the coefficients in all fractional differential masks is not zero, and the response in the region where the gray value of the image is constant or has little change is not zero, so that the result of image texture detail information is enhanced.

Figure 7.

Improved fractional orders of 8 symmetrical directions.

5. Realization of 3D image defect adaptive detection of nano CT

Based on the isotropy of image data, a three-dimensional data processing model is established to process the three-dimensional image sequence of nano CT. Because the interval between layers is very small, the continuity of defect change in casting is strong, and the change of defect target area information in slice is relatively small between two adjacent layers. The region matching algorithm or moment matching algorithm in the target matching can achieve the target matching quickly without searching the matching window. However, for the irregular small target region, the position of the defect target region in the adjacent two initial marker images is not much different, and the shape is not good. It’s also very close. Therefore, a relatively simple method of 3D connected region extraction instead of complex region matching technology is suitable for the processing of 3D image sequence of nano CT. Secondly, there is a big difference between the defect target entity and the video target in the 3D image sequence of nano CT: The defect target has the problem of entity bifurcation. Due to some factors in the manufacturing process, the defect information is bifurcated, or the whole target is divided into several parts due to the segmentation process, so the black dashed circle in the slice image contains two defect targets, while the corresponding area in the slice image contains only one target defect. It is difficult to deal with the target entity with bifurcation by using the region matching technology in the process of multi-target tracking. However, by using the connected region extraction technology, the connected three-dimensional region of bifurcation can be extracted and the data of bifurcation can be marked uniformly.

Based on the above analysis, according to the characteristics of nano CT 3D image sequence and the requirements of defect extraction, and comparing the characteristics of image time series processing, a 3D segmentation and calibration method of multi-target image sequence is proposed. The flow chart is as in Fig. 8.

Figure 8.

3D segmentation process of multi-object image sequence.

The defect extraction process is to mark the defect target area in slice image by using the continuity of information between layers through the initial image sequence (IMIS). Therefore, when marking each layer of image, it is necessary to read in the information of the first and second layers of image. In this paper, we use the connectivity algorithm to mark the connectivity regions. In the calculation process of each layer, since the 8 connectivity regions of the 2D slice image have been extracted, we only need to determine the relationship between the different labeling regions of the current layer slice image and the corresponding regions of the upper and lower layers slice image, as shown in Fig. 9, which is the neighborhood relationship between the pixels of the current k-layer image and the k $-$ 1 layer slice image.

Figure 9.

Relationship between corresponding regions of the image.

The graph on the right is the neighborhood graph of the slice image of k $+$ 1 layer (next layer) corresponding to the current image pixel. The problem of the connected region can be reduced to the relationship between 8 connected regions in the current slice and two 9 connected regions in the upper and lower slices.

Using the above multi-target tracking algorithm to solve the problem of target trajectory bifurcation, the 3D defect target in the initial image sequence (IMIS) can be marked as a 3D defect entity. However, there are many false noise defect targets in the three-dimensional image of nano CT. If the target area is a small target area that only exists in a single layer, as shown in Fig. 9, the noise area will be deleted during the marking process. For more false defect areas that only exist in several layers, the characteristics of the noise area need to be used for detection, as shown in Fig. 10.

Figure 10.

Image slice detection correlation.

As shown in Fig. 10, the black area represents the defect area, and the white area represents the entity. For the current slice image, there are five defect areas, including noise target area and real defect target area. Among them, region 1, 2, 3 and A, B, C of the upper layer region have corresponding neighborhood relations. Region 1, 2, 3 can be remarked respectively, and the mark values of all pixels in these three regions can be changed to corresponding A, B, C. The pixel of region 4 has no corresponding connected neighborhood for the slice image of the upper layer and the lower layer, so the target region only exists in one of the layers is regarded as a false defect. The target area D in the graph has no corresponding neighborhood area in the single front slice, so the defect target track numbered D ends in the graph. However, the target area in the image, although there is no corresponding target area in the corresponding slice image of the previous layer, there is a corresponding area in the next layer, so the area is the starting information of a new defect target, so it should be re marked to realize the adaptive defect detection of 3D image of nano CT.

6. Experimental study

Under different interference conditions, the adaptive defect detection method based on wavelet decomposition and the traditional detection method are tested. The detection results require multi perspective analysis. According to the trend of different wires, the experimental results are observed and summarized.

Figure 11.

Wire images from different perspectives.

Figure 12.

Experimental results.

6.1 Experimental objects

According to PCB wiring rules, the direction of wire mainly includes four cases: horizontal direction, vertical direction, 45 degree angle with horizontal direction and 135 degree angle with horizontal direction. Most of the wires are used to transmit working signals, and their width is basically the same; Generally, there will be one or two layers of physical circuit layer to distribute power lines and ground wires, which are wider than the general wire width; There are other signal wires whose width may be different when they need special consideration, so that the types of wire width of the whole PCB are not many, and the most common is two types (it is a signal line, a power line and a ground line located in a specific circuit layer). The common wire width is 4-12mi1 C lmil, which is 25.4 $\mu$ m. Due to the routing requirements, a wire may include multiple segments, and there is a certain angle between adjacent segments (generally 45 degrees or 135 degrees, rarely 90 degrees), but the width and direction of each segment will remain the same. Because the wire reflects the connection relationship between the pins of each component, each wire area is independent of each other except for the same connection relationship.

6.2 Interference settings

In the experiment, some interferences will be set up, which can be divided into two categories: One is the interferences located in the same physical circuit layer with the conductor, and the other is the interlaminar interferences not belonging to the same physical circuit layer with the conductor. For the first kind of interference, the original drawing is interfered by structural elements smaller than the wire size. The second kind of interference is that the background and the target cannot be separated due to the insufficient contrast of the image, and part of the substrate is also divided into metal substances. The final result may be that the wires of different layers that should not be connected are connected through these substrates that are wrongly divided into metal. The wire image under different interference is shown in Fig. 11.

6.3 Experimental result

In order to better verify the effectiveness of the proposed adaptive detection method based on nano CT 3D image defects, the image filtering algorithm based on adaptive mean in reference [8] is selected for comparative experiments, and the experimental results are shown in Fig. 12.

It can be seen from the test results that the self-adaptive detection method of 3D image defects of nano CT designed in this paper based on wavelet decomposition has obvious broken marks in the transformation process of different perspectives, and has not been affected by the test interference. However, the image defect adaptive detection method in reference [8] has different degrees of influence from different angles, which is a bit fuzzy. At present, it is not clear that due to the traces of image confusion, some images have data appointment, and the recognition of breakpoints is not high. Therefore, the adaptive defect detection method based on wavelet decomposition designed in this paper is more effective. In order to facilitate the accuracy of the test results, the experimental platform Jing data is collected to the storage database for unified identification of the data, and the results are displayed in the same coordinate diagram, as shown in Fig. 13.

Figure 13.

Experimental results of accuracy.

From the observation chart, it can be seen that with the deepening of interference, both methods are affected by different degrees. However, from the point of view of curve change mode, the initial curve slope of this design method is larger than that of reference [8], which indicates that the influence degree is small. It can be seen that this method is more anti-interference and has higher application value.

7. Conclusions

In this paper, the improved white adaptive edge detection algorithm based on wavelet transform is applied to medical images. It retains the advantages of the original algorithm, such as strong anti noise ability, high edge positioning accuracy, good edge continuity, etc. at the same time, it also has a certain ability to detect weak edges. It can detect most of the real edges of medical images, and also has a good ability to suppress false edges. It plays an important role in the auxiliary diagnosis and treatment of the disease. In this paper, based on the prominent dislocations in 3D image detection of nano CT under different interferences, an adaptive defect detection method of 3D image of nano CT based on wavelet decomposition is proposed. This paper focuses on the wavelet decomposition method, taking into account the gray distribution of the image as a whole and the combination with local area information, so as to enhance the resistance to interference under different interference conditions. Through the experimental results, it can be seen that the existing traditional methods are effectively solved.

References

Xiao

and You

S.H.

, Denoising method of engine surface defect image based on wavelet transform, Surface Technology 47(12) (2018), 328–333.

Isah

M.A.

and Kulahcl

M.A.

, A study on null cartan curve in Minkowski 3-space, Applied Mathematics and Nonlinear Sciences 5(1) (2020), 413–424.

Shi

L.Y.

, Method of packaging printing defects detection based on machine vision and image processing, Bulletin of Science and Technology 34(10) (2018), 105–108.

Wang

Yang

H.Y.

Zhao

G.B.

J.X.

Fan

X.D.

and Zhao

G.B.

, Creaform 3D laser scanning technology applied to the detection of pipeline defects, Oil & Gas Storage and Transportation 38(4) (2019), 463–466.

and Yang

, Solution to chance constrained programming problem in swap trailer transport organisation based on improved simulated annealing algorithm, Applied Mathematics and Nonlinear Sciences 5(1) (2020), 47–53.

Zhang

Y.M.

and Xu

B.G.

, Objective evaluation of fabric fuzzing and pilling based on three-dimensional surface image, Cotton Textile Technology 46(1) (2018), 28–33.

X.R.

Ding

J.P.

and Li

J.P.

, Tire impurity defect detection based on morphology and projection histogram, Journal of University of Science and Technology of China 49(1) (2019), 49–54.

Zhang

Z.H.

Yin

X.Z.

and Ren

, Image filtering algorithm of subway tunnel cracks based on adaptive mean value, Journal of Chongqing Jiaotong University (Natural Science Edition) Of China 38(6) (2019), 1–5.

Swetha

and Begum

M.A.

, Design of high speed, area optimized and low power arithmetic and logic unit, Advances in Industrial Engineering and Management 6(1) (2017), 26–31.

10.

Gong

and Gan

, Research on laser 3D image reconstruction based on big data analysis technology, Laser Journal 40(6) (2019), 83–87.

11.

Lan

and Fang

Z.Y.

, Image recognition of steel plate defects based on a 3D gray matrix, Journal of Image and Graphics 24(6) (2019), 859–869.

12.

Yang

X.Y.

Hao

L.T.

Wang

Y.Q.

and Gulzira

, Research progress and trend analysis of advanced 3D Imaging lidar technology, Laser Journal 40(5) (2019), 1–9.

13.

Nath

Chakravarty

A.K.

Ghosh

and Sarkar

S.K.

, FPGA placement optimization using firefly algorithm, Advances in Industrial Engineering and Management 6(2) (2017), 97–102.

14.

Zhou

X.C.

Z.K.

and Shi

L.F.

, Image denoising algorithm based on wavelet packet and partial differential equations, Journal of Electronic Measurement and Instrumentation 32(7) (2018), 61–67.

15.

Mao

T.Q.

Liu

Huang

and Zhao

Y.J.

, Change detection of SAR images using dyadic wavelet enhancement and edge local information FCM, Journal of Signal Processing 34(1) (2018), 54–61.

16.

Ding

and Li

, Project team incentive of the power grid enterprise in China, Advances in Industrial Engineering and Management 8(2) (2019), 82–85.

17.

Wang

Zhang

Y.H.

and Chen

, Image crack detection with multi-scale down-sampled normalized cut, Chinese Journal of Scientific Instrument 38(11) (2017), 2788–2796.

18.

Y.F.

, Character recognition simulation of tilting license plate image based on wavelet analysis, Computer Simulation 36(5) (2019), 190–193.

19.

Fan

and Liu

, Research on steel ball defect detection technology based on machine vision and modified morphological edge detection algorithms, Tool Engineering 53(9) (2019), 102–106.

20.

Z.S.

Y.F.

and Zhu

Y.J.

, Review of defects extraction and recognition methods for welding joints based on X-ray image, Hot Working Technology 47(19) (2018), 6–10.

Adaptive defect detection of 3D image of nano CT based on wavelet decomposition

Abstract

Keywords

1. Introduction

2. Image preprocessing

2.1 Noise analysis of 3D image of nano CT

2.2 Image sequence hybrid filtering

Table 1 Filtering performance parameters

3.1 3D image fusion rules of nano CT based on wavelet decomposition

3.2 Wavebase selection of wavelet decomposition

6.2 Interference settings

6.3 Experimental result

References

Table 1
Filtering performance parameters