Open AccessArticle

Processing of Eddy Current Infrared Thermography and Magneto-Optical Imaging for Detecting Laser Welding Defects

Pengyu Gao

¹,

Xin Yan

¹,

Jinpeng He

¹,

Haojun Yang

¹,

Xindu Chen

¹ and

Xiangdong Gao

^1,2,*

Guangdong Provincial Welding Engineering Technology Research Center, Guangdong University of Technology, No. 100 West Waihuan Road, Higher Education Mega Center, Panyu District, Guangzhou 510006, China

Guangzhou Zhengtian Technology Company, Guangzhou 510006, China

Author to whom correspondence should be addressed.

Metals 2025, 15(2), 119; https://doi.org/10.3390/met15020119

Submission received: 18 November 2024 / Revised: 15 January 2025 / Accepted: 22 January 2025 / Published: 25 January 2025

(This article belongs to the Special Issue Editorial Board Members’ Collection Series: “Welding and Joining”)

Download

Browse Figures

Figure 1
MO and eddy current IR thermography system for detecting laser welding defects. "> Figure 2
Weld defect detection simulation. (a) Three-dimensional model for defect detection. (b) Magnetic flux density map projected onto the surface of the MO film in the defect detection process. The blue box, green box, and red line in (c,d) carry the same significance, all representing the top view of the surface of the workpiece under examination. The blue box delineates the outer contour of a defect with a depth of 1 mm, while the green box indicates a defect with a depth of 0.5 mm. The red line maps the magnetic flux density magnitude values to the corresponding position in (e), implying that the horizontal coordinate in (e) corresponds to the position of the red line along the x-axis of the global coordinate system, with the vertical coordinate representing the magnetic flux density magnitude at that specific location. "> Figure 3
Typical sample image and its processing method. (a) Defect-free sample, (b) burn-out sample, (c) crack sample, (d) incomplete fusion sample, (e) weld bump sample, (f) pit sample. "> Figure 4
The process of weld bump edge extraction by Prewitt operator, which corresponds to <a href="#metals-15-00119-f003" class="html-fig">Figure 3</a>e. Symbol * denotes the convolution operation. "> Figure 5
Experimental system of weld defect ECT detection. (a) Diagram of experimental device, (b) defect detection principle schematic diagram in ECT, (c) welding spot cross-sectional image. "> Figure 6
Laser spot weldment. (a) Laser spot welding specimen image, (b) Welding spot image (red box) collected on microscope, (c) Welding spot image (red box) collected on infrared thermal imager. "> Figure 7
Grayscale curves of the middle column from the magneto-optical image of the crack. "> Figure 8
Comparison of the grayscale curves from the middle column of the crack and defect-free images. "> Figure 9
Visual comparison of samples A1, A2, A3, and A4 before and after denoising. "> Figure 10
Bar graph of BIQI values for each sample before and after denoising of IR thermogram. "> Figure 11
Bar graph of NIQE values for each sample before and after denoising of IR thermogram. "> Figure 12
Bar graph of PSNR values for each sample of IR thermogram. ">

Versions Notes

Abstract

Infrared (IR) magneto-optical (MO) bi-imaging is an innovative method for detecting weld defects, and it is important to process both IR thermography and MO imaging characteristics of weld defects. IR thermography and MO imaging can not only run simultaneously but can also run separately in special welding processes. This paper studies the sensing processing of eddy current IR thermography and MO imaging for detecting weld defects of laser spot welding and butt joint laser welding, respectively. To address the issues of high-level noise and low contrast in eddy current IR detection thermal images interfering with defect detection and recognition, a method based on least squares and Gaussian-adaptive bilateral filtering is proposed for denoising eddy current IR detection thermal images of laser spot welding cracks and improving the quality of eddy current IR detection thermal images. Meanwhile, the image gradient is processed by Gaussian-adaptive bilateral filtering, and then the filter is embedded in the least squares model to smooth and denoise the image while preserving defect information. Additionally, MO imaging for butt joint laser welding defects is researched. For the acquired MO images of welding cracks, pits, incomplete fusions, burn-outs, and weld bumps, the MO image processing method that includes median filtering, histogram equalization, and Wiener filtering was used, which could eliminate the noise in an image, enhance its contrast, and highlight the weld defect features. The experimental results show that the proposed image processing method can eliminate most of the noise while retaining the weld defect features, and the contrast between the welding defect area and the normal area is greatly improved. The denoising effect using the Natural Image Quality Evaluator (NIQE) and the Blind Image Quality Index (BIQI) has been evaluated, further demonstrating the effectiveness of the proposed method. The differences among weld defects could be obtained by analyzing the gray values of the weld defect MO images, which reflect the weld defect information. The MO imaging method can be used to investigate the magnetic distribution characteristics of welding defects, and its effectiveness has been verified by detecting various butt joint laser welding weldments.

Keywords:

laser welding; weld defects; detection; magneto-optical imaging; eddy current infrared thermography; processing

1. Introduction

In the complex laser welding process, welding quality is determined by many factors such as laser power, welding speed, defocus amount, and so on. Due to uncertain environmental factors, welding defects are often generated during laser welding. Welding defect detection is an essential part of the manufacturing process, which directly affects the safety of welding products [1,2]. Weld crack is one of the most serious defects in laser welding, as it is prone to expand under stress and ultimately lead to welding failure. Even tiny cracks can cause significant damage and are often difficult to detect visually [3]. Several conventional techniques are available for welding crack detection, including radiographic testing (RT) [4], ultrasonic testing (UT) [5], and eddy current testing (ECT) [6], but each method has some limitations [4]. To better ensure welding quality and ensure industrial operational safety, there is an urgent need for a more suitable non-destructive testing (NDT) technology.

An innovative method of IR MO bi-imaging has been developed, which includes both MO imaging and infrared imaging techniques [7]. MO imaging is a relatively novel non-destructive testing method [8]. The preparation of MO sensor thin films and the distribution of magnetic fields was studied under direct current excitation [9]. The detection of rail crack defects could be achieved using MO imaging methods [10]. A quantitative measurement method for measuring 3D magnetic field vectors has been proposed using MO imaging [11]. The MO imaging method can be applied to analyze the magnetic distribution and microstructure of weld defects in automatic inspection. As an NDT technique, IR thermography has many advantages such as non-contact; higher detection speed, resolution, and sensitivity; detectability of internal defects; and real-time measurement over a large inspection area [12]. In contrast to visible images, IR imaging can distinguish the target from the background based on differences in thermal radiation and is not easily affected by illumination and environmental light conditions. Previous studies have been conducted on the detection of weld seams and metal cracks using IR imaging techniques [13,14]. Additionally, there are passive IR thermography and active IR thermography techniques [15]. Passive IR thermography is used to measure the temperature difference between the target material and the surrounding environment under different environmental temperature conditions. Active IR thermography uses various external heating sources to generate heat in the material, and the heating distribution can be captured by an IR thermal imager. Eddy current pulsed thermography just uses eddy currents as a heating source, which combines the advantages of eddy current and IR thermography. An eddy current is induced to heat conductive materials, and an IR thermal imager is used to detect the thermal radiation waves emitted from the interface boundary.

Eddy current pulsed thermography has been proven to be effective for crack detection in welding [16]. However, the IR imaging method is different from that of visible images, and the energy of IR radiation is often low and uneven, which could result in information loss of the IR images. Moreover, during the acquisition of IR images, noise interference, unclear defect features caused by uneven heating, random disturbances from the external environment, and the inherent characteristics of IR thermal imagers lead to a low signal-to-noise ratio and low contrast, which can lead to the degradation of image information, affecting image quality and visualization effects and interfering with weld defect recognition and positioning. Therefore, improving the quality of IR images and reducing noise interference is crucial for effectively acquiring and utilizing eddy current IR detection thermal image information for welding defect detection and positioning [17].

Due to the influence of factors such as high temperature and gases during laser welding, eddy current IR detection thermal images of welding defects are prone to noise and may have much more noise compared with other IR images in non-welding environments. Additionally, there are often specific textures present in the weld points that also affect the noise level. Therefore, it is important to study an image denoising method that can achieve good denoising in eddy current IR detection thermal images of welding defects. Bilateral filtering (BLF) is a method of local edge-preserving smoothing, which can effectively remove noise and smooth images by considering the similarity between pixels and their spatial distances while preserving defect edges and weld point structural information, making defects appear more clearly. Image processing using BLF is well suited to the requirements of defect detection for the acquired defect image; this aids in more accurate defect identification, localization, and analysis, further improving the efficiency and accuracy of defect detection [18]. It can be observed that using bilateral filtering is a highly effective method that aligns with the requirements for processing eddy current IR detection thermal images of welding defects. Additionally, some early attempts have focused on reducing the computational burden, so BLF has fast approximations and is an option for real-time or near-real-time image processing, which is advantageous for real-time defect detection and aligns with practical application needs [19]. As a filtering and denoising method, BLF nonlinearly combines nearby image values based on geometric closeness and photometric similarity in grayscale or color [20]. Some researchers have already applied BLF to IR image processing. An adaptive BLF method with an improved convolution kernel for IR image enhancement has been proposed. This method combines edge detection operators with BLF [21]. The input image is divided by the relativity of a Gaussian-adaptive BLF into a base layer and a detail layer. The detail layer is multiplied by the proposed weight coefficient, while the base layer is processed by histogram projection. Finally, the detail layer and base layer are combined for further processing [22]. The drawback of BLF is its tendency to produce gradient reversal and halo, which can adversely affect the appearance of the images [23].

These BLF-based methods have achieved good results in IR imaging. However, the complexity of noise and texture increases the processing difficulty for eddy current IR detection thermal images of welding defects, and traditional methods may not achieve optimal denoising effects. Therefore, a processing method that can exhibit good denoising performance on eddy current IR detection thermal images of welding defects is needed. In this paper, butt joint laser welding and laser spot welding defects were tested, which verified the effectiveness and universality of IR and MO sensing for detecting weld defects. The enhanced Prewitt operator and modified MO and IR image feature enhancement (MIIE) algorithm have been applied to meet the purpose of only highlighting the welding defect features. The denoising method of Gaussian-adaptive BLF embedded in a least squares model is applied to the eddy current IR detection thermal images of laser spot welding cracks and verified the effectiveness of this method. Additionally, an MO and IR image feature enhancement algorithm was employed for image enhancement, which could enhance the feature representation capability of the images and reduce noise.

The MO excitation device employed in the experiments is a U-shaped electromagnet, and the magnetic field that it generates is unidirectional concerning the defect’s position, leading to a more pronounced leakage magnetic phenomenon at the edges perpendicular to the magnetic field. This characteristic is optimally mapped by an MO sensor. Essentially, the MO sensor excels in characterizing linear defects compared to arcuate or circular ones, whereas the defects caused by laser seam welding are primarily rectangular. In contrast, spot welding may result in either linear or arcuate defects. Therefore, an MO sensor is employed for seam welding detection, while the infrared detection capabilities are relatively comprehensive, with an IR sensor used to supplement image acquisition. Furthermore, the information on the depth of defects could be ascertained by the MO image characteristics.

2. Experimental Equipment and Methods

The experimental weldments in this study are derived from laser spot welding and butt joint laser welding, with the experimental system shown in Figure 1. For detecting weld defects, simulation experiments and actual process tests are carried out.

2.1. Experimental Setup and Processing for MO Images

An MO sensor (Magview camera, Matesy, Jena, Germany) was used to capture the weldment spot image, whose working principle is based upon the Faraday MO rotation effect [8,9]. Experimental samples in this study are derived from laser spot welding and butt joint laser welding, with the experimental system shown in Figure 1. The devices such as motion mechanism, inductive heating system, and excitation control system are integrated and manufactured by Zhengtian Company (Guangzhou, China). A MO imaging detection software (V1.0) was used to process the MO images. For detecting weld defects, simulation experiments and actual process tests are carried out. A simulation of the magnetic field conditions during the detection of welding defects in steel plates using an MO sensor has been conducted to validate the effectiveness of the detection results. The simulated defect three-dimensional model comprises a combination of defects of dimensions 3 mm (length) × 1 mm (width) × 1 mm (depth) and 3 mm (length) × 1 mm (width) × 0.5 mm (depth). The MO film of the sensor is situated 0.5 mm above the surface of the steel plate. The simulation results and the relative position of the model are illustrated in Figure 2. In Figure 2b, it is evident that there is a sudden increase in the magnetic flux density at the defect locations, with a more pronounced increase in the region at a depth of 1 mm, which is related to the principle of magnetic leakage. When encountering a leakage magnetic field caused by weld defects within the MO film, the light in the MO sensor undergoes deflection and is detected through a polarizing filter, captured by a camera, and an MO image of the defect is generated.

The angle of light deflection

θ

can be calculated using the Faraday effect, as shown in Formula (1).

B

is the applied magnetic induction strength;

V

is the Faraday rotation constant of the material, also known as the Verdet constant, the value of which depends on the material properties and the wavelength of light; and

L

is the thickness of the MO film. In other words, changes in the magnetic field will be reflected in the MO image, and the sudden increase in magnetic flux at the welding defect location will be mapped in an MO image.

θ = V × B × L

(1)

Simulation software (COMSOL V6.1) is employed to segment the various components of the MO detection scenario and assign them distinct magnetization models. The models for air, copper coils, and the magnetic core of the electromagnet were configured as a relative permeability model, with its constitutive relationship outlined in Formula (2), where B denotes the magnetic induction strength,

μ_{0}

represents the permeability of free space,

μ_{0}

signifies the relative permeability, and H indicates the magnetic field strength. The magnetization model for the soft magnetic specimens was defined according to the B-H curve, with the constitutive relationship illustrating the correlation between the magnetic induction strength and magnetic field strength as depicted in Formula (3). The occurrence of magnetic leakage primarily stems from the difference in relative permeability between air and the soft magnetic materials.

B = μ_{0} μ_{r} H

(2)

|H| = f (|B|)

(3)

A typical weldment sample image and its processing method are shown in Figure 3, and the sampling signal is used by an MO imaging sensor. From the sampled images in the second column of Figure 3 and the corresponding three-dimensional grayscale feature map, it can be observed that, in addition to the main features, there is also noise from various interference factors. To enhance the feature representation capability of the images and reduce noise, the MIIE algorithm was employed for image enhancement, with the specific process outlined in Formula (4). The MO image processing segment utilized in this experiment includes median filtering, histogram equalization, and Wiener filtering. Median filtering serves to eliminate scratch noise within the images, while histogram equalization enhances the contrast of the images. Following the contrast enhancement, Gaussian noise in the finer details of the image is accentuated. Consequently, Wiener filtering is employed to mitigate the Gaussian noise present in the images.

O_{1} [i, j] = \{\begin{matrix} m e d i a n (k e r n a l_{1} (i, j) i f s t d (k e r n a l_{1} (i, j)) > T_{1} \\ I [i, j] otherwise \end{matrix}, O_{2} [i, j] = \sum_{t = 0}^{k} p_{r} (t) \times (n o r m a l i z e (O_{1} [i, j])) \times 255, I_{o u t p u t} (i, j) = f_{w i e n e r} \{O_{2} [i, j]\}

(4)

where i, j are the pixel position at the i-th row and j-th column in an image.

$s t d ()$ is the standard deviation calculation.
$k e r n a l_{1} (i, j)$ is the neighborhood of the input pixel (i, j).
$I [i, j]$ is the grayscale value of the pixel at the i-th row and j-th column of the original image.
$m e d i a n ()$ is the median calculation.
$T_{1}$ is the set threshold.
$n o r m a l i z e ()$ is the normalization calculation.
k is the index of a specific grayscale level.
$p_{r} (t)$ is the probability of grayscale value t appearing in the image.
$f_{w i e n e r} {}$ is the Wiener filtering.
$I_{o u t p u t} (i, j)$ is the image gray value of the output image at position $(i, j)$ .

Figure 3. Typical sample image and its processing method. (a) Defect-free sample, (b) burn-out sample, (c) crack sample, (d) incomplete fusion sample, (e) weld bump sample, (f) pit sample.

Since the enhanced images exhibit significant gradient differences at the features, while the overall trend of the gray level changes in the MO images is smooth, the Prewitt operator was considered to strengthen the edge features of the images, serving as auxiliary features to help in visualizing the defects, as detailed in Formula 5.

K_{x} = [\begin{matrix} 1 & 1 & 1 \\ 0 & 0 & 0 \\ - 1 & - 1 & - 1 \end{matrix}], K_{y} = [\begin{matrix} - 1 & 0 & 1 \\ - 1 & 0 & 1 \\ - 1 & 0 & 1 \end{matrix}], I_{x} = I * K_{x}, I_{y} = I * K_{y},

O u t p u t = \sqrt{I_{x}^{2} + I_{y}^{2}}

(5)

where K_x represents the Prewitt operator in the horizontal direction, which is capable of extracting gradients in the vertical direction, thereby detecting horizontal edges. Conversely, K_y denotes the Prewitt operator in the vertical direction, which extracts gradients in the horizontal direction, thus enabling the detection of vertical edges. I represents the input image, while

*

denotes the convolution operation. By applying the convolution kernel

K_{x}

to the image

I

, one can obtain the gradient of the images in the horizontal direction

I_{x}

I_{y}

represents the gradient of the image in the vertical direction. Ultimately, by combining these two operators through a weighted approach, a unified edge map can be obtained, as shown in Figure 4.

2.2. Experiments for Eddy Current IR Thermography

The experimental setup for detecting weld defects by eddy current IR thermography is shown in Figure 5a. It mainly consists of an inductive heating system with an inductive coil, a cooling system, and an IR thermal imager made by IRay (Yantai, China). The inductive heating system is capable of heating the weldments with an output power of 2 kW and a frequency of 50 kHz. The cooling system provides overheating protection for the inductive heating system. The IR thermal imager records the temperature distribution and variations on the surface of the weldments and outputs the heat information in the form of a thermal image. The main parameters of the IR thermal imager are as follows: the temperature measurement range is 0–200 °C, the thermal sensitivity of the IR imager is less than 40 mk, the spatial resolution is 640 × 512, and the measurement wavelength is 8–14 μm. The object distance of the camera lens in the experimental data acquisition equipment is set at 79.3 mm, which guarantees that both the heating of the steel plate and the detection tasks can be performed simultaneously without interference from the magnetic field of the electromagnetic induction coil in the heating process.

The schematic diagram in Figure 5b illustrates the principle of eddy current thermography (ECT) defect detection. The process of ECT defect detection mainly involves induction heating and heat conduction such that when an alternating current is passed through the coil, the conductor in the vicinity of the coil will generate induction eddy current. When defects are present in a conductor, the distribution of the eddy currents is disturbed, leading to uneven heat distribution in the conductor. By utilizing IR thermal imaging to capture the temperature distribution of IR radiation, we can convert this information into IR thermal radiation images. This allows for the observation of the size, location, and type of weld defects. The cross-sectional view of the weld point is illustrated in Figure 5c.

Laser spot welding is a common method of welding metal materials that is widely used in various fields such as automotive, aerospace, medical equipment, electronics, and household appliances. It is well known that the detection of welding defects such as cracks is of great importance to ensure the welding quality [24,25]. Here, lap welding of two 65 Mn steel plates with dimensions of 200 mm × 100 mm × 1.5 mm is used as the test weldments, and laser spot welding cracks are detected by the ECT method. In the experiment, the laser power is 2 kW, and the welding time is 500 ms. Eddy current IR detection thermal images of the crack specimen are obtained through the experimental setup. Figure 6 shows an example of the laser welding spot.

To decrease the noise influence on IR thermal images of laser spots, the BLF model, Gaussian-adaptive bilateral filtering (GABF) model [26], and least squares (LS) embedded with Gaussian-adaptive bilateral filter are applied. GABF improves the processing effectiveness of BLF by improving the Gaussian range kernel; meanwhile, LS reduces the computational workload. In order to balance the smoothing and denoising effects with operational efficiency, GABF can be embedded within the LS framework. The main steps of this approach involve first applying GABF to smooth the gradient of the IR image and then integrating the smoothed gradient into the LS framework. This method preserves image details while enhancing the smoothing and denoising effects and optimizing computational efficiency. The model is given as Formula (6).

\min_{u} \sum_{i} ({(u_{i} - I_{i})}^{2} + λ \sum_{* \in \{x, y\}} {(\nabla u_{*, i} - (f_{GABF} (\nabla I_{*}, \bar{g}))_{i})}^{2})

(6)

where I is an input infrared image, u is an output image, and i and j denote pixel positions. λ is responsible for adjusting the balance between the two terms, and increasing the value of λ gradually smoothens the image.

\bar{g}

is a low-pass guidance image.

\nabla I_{x}

and

\nabla I_{y}

represent the gradients of the input infrared image I at the pixel point along the x-axis and y-axis, respectively, while

f_{GABF} (\nabla I_{x}, \bar{g})

and

f_{GABF} (\nabla I_{y}, \bar{g})

represent the results of smoothing the gradients using GABF, respectively. In summary, the denoising method for eddy current infrared detection thermal images of welding defects can be divided into two steps such as smoothing the gradients along the x-axis and y-axis of the input image using GABF and solving Formula (6) to obtain the smoothed image.

3. Experimental Results and Discussion

3.1. Extraction of Weld Defect Characteristics of MO Images

As shown in the fourth column of Figure 3, the images processed with the Prewitt operator can roughly reflect the general morphology of various features. From the gradient map and the 3D gradient map, it can be observed that the characteristic gradients of different defects vary significantly. The gradient peak for the defect-free sample is the lowest and most smooth. The gradient peak for the burn-out defect is higher and wider. The weld bump that occurs alongside the burn-out exhibits a more fluctuating gradient than that of the burn-out defect. The gradient peak for the crack is the highest and most uniform, and the gradient peak for incomplete fusion is relatively high and shows noticeable changes at the defect site. The characteristics of the pit are similar to those of incomplete fusion, as both share structural similarities.

Focusing on the MO image of the crack, the grayscale values from the middle column of the image were extracted, as shown in Figure 7. It can be observed that after MIIE processing, the features of the crack become more pronounced. Furthermore, following the application of the Prewitt edge extraction method, the morphology of the crack can be effectively constructed.

Apart from cracks and defect-free images, other types of defects may appear at different horizontal positions within the image. Therefore, only the grayscale values from the middle column of the crack and defect-free images can be extracted for comparative analysis, as shown in Figure 8. It can be observed that at the location of the crack, the grayscale value rapidly increases from 60 to 230, whereas in the defect-free area, the grayscale value rises more gradually from 30 to 165. This indicates that the most significant difference between the crack and the defect-free region is the abrupt change in grayscale value at the crack location. Furthermore, from the grayscale curve of the image processed by the Prewitt operator, it can be observed that the maximum gradient of the crack is 100, whereas the maximum gradient of the defect-free area is only 40.

3.2. Denoising of Thermal Images of Weld Defects

The peak signal-to-noise ratio (PSNR) is defined based on the Mean Squared Error (MSE) and is used to quantify the difference between the original image and the distorted image. A higher PSNR value indicates less distortion and better image quality. The calculation of the PSNR begins with determining the pixel values of both the original and the processed images. Subsequently, the errors between the corresponding pixel points of the two images are computed, followed by calculating the sum of the squared errors for all pixel points. Finally, this squared sum is divided by the total number of pixels in the image to obtain the MSE. In the context of image enhancement, the PSNR serves as a measure of similarity between the enhanced image and the original image. A higher PSNR value signifies that the enhanced image is more similar to the original image, indicating lower levels of distortion; typically, a PSNR value of 40 suggests excellent quality.

Due to the fact that ideal noise-free eddy current IR detection thermal images of welding defects are unknown, it is necessary to use no-reference image quality evaluation standards to evaluate the overall denoising effect. No-reference image quality evaluation models allow for evaluating the quality of an image without any reference image, solely based on its own statistical characteristics. In this paper, we primarily utilize no-reference image quality evaluation standards to evaluate image quality, which include the Blind Image Quality Index (BIQI) model and the Natural Image Quality Evaluator (NIQE) model [27,28]. The BIQI model is based on the natural scene statistic, where the image is decomposed using the wavelet transform over three scales and orientations and uses the generalized Gaussian distribution to fit the distribution of wavelet coefficients in each sub-band, and the fitted parameters are used as features to perceive changes in the quality of the image. Then, support vector regression is used to predict the quality of the image and calculate the objective image quality score. The smaller the BIQI is, the better the image quality is.

The NIQE metric establishes a series of features that can characterize the quality of an image and uses a Gaussian model to describe them. This means that it fits the distribution of features using a multivariate Gaussian distribution on a natural image set, with the distribution parameters serving as features for the images; then, for a given image to be evaluated, the same operation is performed to extract the distribution parameters, and the quality of the image is then measured by analyzing the distance between the distribution parameters of the natural images and the distribution parameters of the image to be evaluated. This evaluation metric is more in line with the judgment of the human visual system as it is based on the statistical characteristics of natural images. The final expression of the NIQE metric is as in Formula (7).

D (v_{1}, v_{2}, \sum_{1}, \sum_{2}) = \sqrt{({(v_{1} - v_{2})}^{T} {(\frac{\sum_{1} + \sum_{2}}{2})}^{- 1} (v_{1} - v_{2}))}

(7)

where

v_{1}

v_{2}

and

\sum_{1}

\sum_{2}

represent the mean vector and covariance matrix of the multivariate Gaussian model for the natural image and the image to be evaluated, respectively. It can be observed that a smaller NIQE value indicates that the evaluated image is visually closer to a natural image, resulting in a more natural and higher-quality image. The training of the NIQE algorithm does not depend on any subjective quality scores, and the algorithm has a low complexity for image quality calculation, making it suitable for real-time systems.

Figure 9 shows the denoised results using the proposed denoising algorithm on eddy current IR detection thermal images of four different laser spot weld crack specimens A1, A2, A3, and A4, which are shown in Figure 9a, Figure 9b, Figure 9c, and Figure 9d, respectively, with a Gaussian blur radius of 25, a geometric proximity coefficient of 4, and an intensity similarity coefficient of 0.005. In each figure, the left side displays the original eddy current IR detection thermal images of welding defects, and the right side shows the zoomed-in thermogram of the welding spots, where the top is the original image and the bottom is the denoised image. The results shown in the figure indicate that, after processing by the described method, the welding spot images appear visually smoother, the contours of the welding spots and the features of the defects are preserved, and the denoising effect is significant with a noticeable reduction in noise. The denoised images exhibit better visual quality, enabling welding spots and defects to be clearly shown. This processing effect is very beneficial for subsequent defect detection algorithms and manual operators in industrial applications, providing more helpful image data that can improve accuracy and reliability in defect analysis.

The Otsu algorithm is an adaptive thresholding image segmentation method, also known as the maximum between-class variance method [29]. It uses the grayscale information of the image to categorize the image information into two classes, target and background, and selects a segmentation threshold with the criterion of it maximizing the variance between the target class and background class. Through segmentation, we can effectively distinguish different elements in the weld image. The ideal segmentation result for welding defect images is to accurately separate the defects, welding spot contours, and background while maintaining clear and accurate boundaries. It is essential to preserve the integrity of each region to enable accurate identification and separation of the defect areas. Regarding the ideal segmentation result before and after denoising, noise is often considered to be an interfering signal that may confuse or mask the defect features within the welding spots, so the goal is to assign noise to the background class and exclude it from the defect and welding spot contours when performing the segmentation. In this paper, the Otsu algorithm is used to segment the eddy current IR detection thermal images of laser spot welding crack specimens. Table 1 presents the Otsu threshold segmentation results of the original images and denoised images for the four samples. It can be observed from the experiments that the image segmentation before denoising shows many small, discrete noise regions in the segmented images, these noise points are classified as the target class, similar to the defect features. However, the denoising process effectively smooths the noise in the images, making it closer to the grayscale values of the background pixels. After segmentation of the processed images, the noise pixels are more accurately classified as background, thus reducing interference in defect identification. A comparison of before and after denoising shows that the internal defect features within the weld spots become more pronounced in the processed images, which further confirms the effectiveness of the denoising.

Additionally, the IR thermograms before and after denoising of the four welding defect samples were evaluated using the BIQI, NIQE, and PSNR as evaluation indicators, and the results of the evaluation are shown in Table 2, Table 3 and Table 4. Figure 10, Figure 11 and Figure 12 present the bar graphs of the BIQI, NIQE, and PSNR values for the four samples before and after denoising, respectively. It is worth noting that lower values of the BIQI and NIQE and higher values of the PSNR indicate better image quality, and it can be seen from the graphs that both evaluation indicator values for the original infrared thermograms of each sample are better than those of the denoised infrared thermograms. According to the data in Table 2, Table 3 and Table 4, the average BIQI value for the original IR thermograms of all samples is 58.6621; at the same time, after denoising, the average BIQI value for all samples is 47.4545, which indicates a 19.1054% improvement in image quality. Similarly, the average NIQE value for the original IR thermograms of all samples is 36.4581. After denoising, the average NIQE value for the original IR thermograms of all samples is 27.8268, resulting in a 23.6746% improvement in image quality. The average PSNR value is 45.8, which exceeds 40, indicating that the denoising effect is very effective. Again, these results show that denoising has a significant effect on improving the image quality of the eddy current IR detection thermal images of cracked weldments.

4. Conclusions

Both MO imaging and eddy current IR thermography are effective for detecting weld defects, but they are susceptible to noise interference, which can reduce the detection accuracy. This paper explores filtering algorithms to enhance image quality and highlight the weld defect features.

The MO imaging method, based on the Faraday rotation effect, detects laser welding defects by analyzing the magnetic distribution. Simulations show that magnetic leakage increases with defect depth, allowing for the determination of defect presence and depth. The MIIE method, combining sliding median transformation, histogram equalization, and Wiener filtering, is used to improve image contrast and reduce noise. The Prewitt operator further enhances edge features, making the weld defects more visible.

For eddy current IR thermography, a denoising method using least squares and Gaussian-adaptive bilateral filtering is proposed. This method effectively removes noise from thermal images of laser spot weld cracks while preserving image information. The algorithm’s effectiveness is verified through comparisons of processed and unprocessed thermograms, Otsu threshold segmentation, and no-reference image quality evaluation. The proposed method performs well in denoising eddy current IR detection thermal images of laser spot weld cracks.

Author Contributions

Conceptualization, X.G.; methodology, P.G., X.Y., J.H., H.Y. and X.C.; validation, P.G. and X.Y.; investigation, P.G., X.Y. and J.H.; resources, X.G.; data curation, P.G., X.Y., J.H. and H.Y.; writing—original draft preparation, P.G., X.Y. and J.H.; writing—review and editing, X.G. and X.C.; visualization, P.G., X.Y. and H.Y.; supervision, X.G. and X.C.; project administration, X.G.; funding acquisition, X.G. All authors have read and agreed to the published version of the manuscript.

Funding

This work was supported in part by the Guangdong Provincial Natural Science Foundation of China (2023A1515012172) and the Guangzhou Municipal Special Fund Project for Scientific and Technological Innovation and Development under Grant (2023B03J1326).

Data Availability Statement

The original contributions presented in this study are included in the article. Further inquiries can be directed to the corresponding author.

Conflicts of Interest

Author Xiangdong Gao was employed by Guangzhou Zhengtian Technology Company. The remaining authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

References

Zhang, Y.; Gao, X.; You, D.; Zhang, N. Data-driven detection of laser welding defects based on real-time spectrometer signals. IEEE Sens. J. 2019, 19, 9364–9373. [Google Scholar] [CrossRef]
Fan, X.; Gao, X.; Zhang, N.; Ye, G.; Liu, G.; Zhang, Y. Monitoring of 304 austenitic stainless-steel laser-MIG hybrid welding process based on EMD-SVM. J. Manuf. Process. 2022, 73, 736–747. [Google Scholar] [CrossRef]
Broberg, P. Surface crack detection in welds using thermography. NDT E Int. 2013, 57, 69–73. [Google Scholar] [CrossRef]
Yan, Z.; Xu, H.; Huang, P. Multi-scale multi-intensity defect detection in ray image of weld bead. NDT E Int. 2020, 116, 102342. [Google Scholar] [CrossRef]
Munir, N.; Kim, H.; Park, J.; Song, S.; Kang, S. Convolutional neural network for ultrasonic weldment flaw classification in noisy conditions. Ultrasonics 2019, 94, 74–81. [Google Scholar] [CrossRef]
Bellanova, M.; Carboni, M.; Felicetti, R. A feasibility analysis on the application of eddy current testing to the detection of the most detrimental defects in historical metallic tie-rods. NDT E Int. 2019, 104, 34–41. [Google Scholar] [CrossRef]
He, J.; Gao, X.; Yang, H.; Gao, P.; Zhang, Y. Analysis of Image Formation Laws and Enhancement Methods for Weld Seam Defects Based on Infrared and Magneto-Optical Sensor Technology. J. Nondestruct. Eval. 2024, 43, 101. [Google Scholar] [CrossRef]
Polyanskii, A.; Larbalestier, D. Magneto-optical characterization techniques. In Handbook of Superconductivity; CRC Press: Boca Raton, FL, USA, 2022; pp. 371–382. [Google Scholar]
Arakelyan, S.; Galstyan, O.; Lee, H.; Babajanyan, A.; Lee, J.; Friedman, B.; Lee, K. Direct current imaging using a magneto-optical sensor. Sens. Actuators A Phys. 2016, 238, 397–401. [Google Scholar] [CrossRef]
Xu, C.; Xu, G.; He, J.; Cheng, Y.; Dong, W.; Ma, L. Research on rail crack detection technology based on magneto-optical imaging principle. J. Phys. Conf. Ser. 2022, 2196, 012003. [Google Scholar] [CrossRef]
Sakaguchi, H.; Oya, R.; Wada, S.; Matsumura, T.; Saito, H.; Ishibashi, T. 3D magnetic field vector measurement by magneto-optical imaging. J. Magn. Soc. Jpn. 2022, 46, 37–41. [Google Scholar] [CrossRef]
Tong, Z.; Xie, S.; Liu, H.; Zhang, W.; Pei, C.; Li, Y.; Chen, Z.; Uchimoto, T.; Takagi, T. An efficient electromagnetic and thermal modelling of eddy current pulsed thermography for quantitative evaluation of blade fatigue cracks in heavy-duty gas turbines. Mech. Syst. Signal Proc. 2020, 142, 106781. [Google Scholar] [CrossRef]
Xie, J.; Xu, C.; Chen, G.; Huang, W.; Song, G. Automated identification of front/rear surface cracks in ferromagnetic metals based on eddy current pulsed thermography. Infrared Phys. Technol. 2022, 126, 104345. [Google Scholar] [CrossRef]
Liang, X.; Peng, J.; Zhang, X.; Guo, J.; Zhang, Y. Analysis of the RCF crack detection phenomenon based on induction thermography. Appl. Optics 2022, 61, 4809–4816. [Google Scholar] [CrossRef] [PubMed]
Xu, C.; Zhang, W.; Wu, C.; Xie, J.; Yin, X.; Chen, G. An improved method of eddy current pulsed thermography to detect subsurface defects in glass fiber reinforced polymer composites. Compos. Struct. 2020, 242, 112145. [Google Scholar] [CrossRef]
Song, J.; Gao, B.; Woo, W.; Tian, G. Ensemble tensor decomposition for infrared thermography cracks detection system. Infrared Phys. Technol. 2020, 105, 103203. [Google Scholar] [CrossRef]
Shao, Y.; Sun, Y.; Zhao, M.; Chang, Y.; Zheng, Z.; Tian, C.; Zhang, Y. Infrared image stripe noise removing using least squares and gradient domain guided filtering. Infrared Phys. Technol. 2021, 119, 103968. [Google Scholar] [CrossRef]
Tomasi, C.; Manduchi, R. Bilateral filtering for gray and color images. In Proceedings of the Sixth International Conference on Computer Vision (IEEE Cat. No. 98CH36271), Bombay, India, 4–7 January 1998; pp. 839–846. [Google Scholar]
Chaudhury, K.N.; Dabhade, S.D. Fast and provably accurate bilateral filtering. IEEE Trans. Image Process. 2016, 25, 2519–2528. [Google Scholar] [CrossRef] [PubMed]
Goyal, J.B.; Dogra, A.; Agrawal, S.; Sohi, B.S.; Sharma, A. Image denoising review: From classical to state-of-the-art approaches. Inf. Fusion 2020, 55, 220–244. [Google Scholar] [CrossRef]
Lv, H.; Shan, P.; Shi, H.; Zhao, L. An adaptive bilateral filtering method based on improved convolution kernel used for infrared image enhancement. Signal Image Video Process 2022, 16, 2231–2237. [Google Scholar] [CrossRef]
Feng, X.; Pan, Z. Detail enhancement for infrared images based on Relativity of Gaussian-Adaptive Bilateral Filter. OSA Contin. 2021, 4, 2671–2686. [Google Scholar] [CrossRef]
Liu, W.; Zhang, P.; Chen, X.; Shen, C.; Huang, X.; Yang, J. Embedding Bilateral Filter in Least Squares for Efficient Edge-Preserving Image Smoothing. IEEE Trans. Circuits Syst. Video Technol. 2020, 30, 23–35. [Google Scholar] [CrossRef]
Wang, R.; Wang, H.; He, Z.; Zhu, J.; Zuo, H. Weld Net: A lightweight deep learning model for welding defect recognition. Weld World 2024, 68, 2963–2974. [Google Scholar] [CrossRef]
Ma, N.; Gao, X.; Tian, M.; Wang, C.; Zhang, Y.; Gao, P. Magneto-optical imaging of arbitrarily distributed defects in welds under combined magnetic field. Metals 2022, 12, 1055. [Google Scholar] [CrossRef]
Chen, B.H.; Tseng, Y.S.; Yin, J.-L. Gaussian-Adaptive Bilateral Filter. IEEE Signal Process. Lett. 2020, 27, 1670–1674. [Google Scholar] [CrossRef]
Moorthy, A.K.; Bovik, A.C. A Two-Step Framework for Constructing Blind Image Quality Indices. IEEE Signal Process. Lett. 2010, 17, 513–516. [Google Scholar] [CrossRef]
Mittal, A.; Soundararajan, R.; Bovik, A.C. Making a “Completely Blind” Image Quality Analyzer. IEEE Signal Process. Lett. 2013, 20, 209–212. [Google Scholar] [CrossRef]
Otsu, N. A Threshold Selection Method from Gray-Level Histograms. IEEE Trans. Syst. Man Cybern. 1979, 9, 62–66. [Google Scholar] [CrossRef]

Figure 1. MO and eddy current IR thermography system for detecting laser welding defects.

Figure 2. Weld defect detection simulation. (a) Three-dimensional model for defect detection. (b) Magnetic flux density map projected onto the surface of the MO film in the defect detection process. The blue box, green box, and red line in (c,d) carry the same significance, all representing the top view of the surface of the workpiece under examination. The blue box delineates the outer contour of a defect with a depth of 1 mm, while the green box indicates a defect with a depth of 0.5 mm. The red line maps the magnetic flux density magnitude values to the corresponding position in (e), implying that the horizontal coordinate in (e) corresponds to the position of the red line along the x-axis of the global coordinate system, with the vertical coordinate representing the magnetic flux density magnitude at that specific location.

Figure 4. The process of weld bump edge extraction by Prewitt operator, which corresponds to Figure 3e. Symbol * denotes the convolution operation.

Figure 5. Experimental system of weld defect ECT detection. (a) Diagram of experimental device, (b) defect detection principle schematic diagram in ECT, (c) welding spot cross-sectional image.

Figure 6. Laser spot weldment. (a) Laser spot welding specimen image, (b) Welding spot image (red box) collected on microscope, (c) Welding spot image (red box) collected on infrared thermal imager.

Figure 7. Grayscale curves of the middle column from the magneto-optical image of the crack.

Figure 8. Comparison of the grayscale curves from the middle column of the crack and defect-free images.

Figure 9. Visual comparison of samples A1, A2, A3, and A4 before and after denoising.

Figure 10. Bar graph of BIQI values for each sample before and after denoising of IR thermogram.

Figure 11. Bar graph of NIQE values for each sample before and after denoising of IR thermogram.

Figure 12. Bar graph of PSNR values for each sample of IR thermogram.

Table 1. Comparison test results using Otsu threshold segmentation for weld images.

	Defect A1	Defect A2	Defect A3	Defect A4
Original detectedimage
Otsu
Denoised detected image
Otsu

Table 2. BIQI before and after denoising of infrared thermograms of specimens A1–A4.

Specimen Serial Number	Original Image	Denoised Image	Percentage Improvement in Image Quality
A1	83.1304	62.9606	24.2628%
A2	49.4360	48.1192	2.6636%
A3	53.3947	32.8199	38.5334%
A4	48.6874	45.9181	5.6879%
Average	58.6621	47.4545	19.1054%

Table 3. NIQE before and after denoising of infrared thermograms of specimens A1–A4.

Specimen Serial Number	Original Image	Denoised Image	Percentage Improvement in Image Quality
A1	36.5935	20.7790	43.2167%
A2	33.2001	19.9268	39.9797%
A3	19.6289	15.3424	21.8377%
A4	56.4100	55.2588	2.0408%
Average	36.4581	27.8268	23.6746%

Table 4. PSNR of infrared thermograms of specimens A1–A4.

Specimen Serial Number	A1	A2	A3	A4
PSNR	48.4	43.7	50.3	40.9
Average	45.8

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

© 2025 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https://creativecommons.org/licenses/by/4.0/).

Share and Cite

MDPI and ACS Style

Gao, P.; Yan, X.; He, J.; Yang, H.; Chen, X.; Gao, X. Processing of Eddy Current Infrared Thermography and Magneto-Optical Imaging for Detecting Laser Welding Defects. Metals 2025, 15, 119. https://doi.org/10.3390/met15020119

AMA Style

Gao P, Yan X, He J, Yang H, Chen X, Gao X. Processing of Eddy Current Infrared Thermography and Magneto-Optical Imaging for Detecting Laser Welding Defects. Metals. 2025; 15(2):119. https://doi.org/10.3390/met15020119

Chicago/Turabian Style

Gao, Pengyu, Xin Yan, Jinpeng He, Haojun Yang, Xindu Chen, and Xiangdong Gao. 2025. "Processing of Eddy Current Infrared Thermography and Magneto-Optical Imaging for Detecting Laser Welding Defects" Metals 15, no. 2: 119. https://doi.org/10.3390/met15020119

APA Style

Gao, P., Yan, X., He, J., Yang, H., Chen, X., & Gao, X. (2025). Processing of Eddy Current Infrared Thermography and Magneto-Optical Imaging for Detecting Laser Welding Defects. Metals, 15(2), 119. https://doi.org/10.3390/met15020119

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Menu