This thesis focuses on the image simulation of shrinkage cavity and porosity defects in castings, particularly within the framework of an X-ray intelligent inspection system. The primary objective is to develop methods for generating realistic virtual defects that can be used to evaluate and tune defect detection algorithms without the need for extensive real-world image acquisition. My research covers three main aspects: an efficient algorithm for measuring the diameter of irregular defects, the application of partial differential equation (PDE) based diffusion filtering for image denoising, and a novel approach for simulating shrinkage cavity and porosity defects in X-ray images. The proposed simulation method constructs defects based on spatial characteristics of the casting image, incorporating randomness and accuracy in shape, gray level, and size. Experimental results demonstrate that the simulated defects closely resemble real defects and meet the required precision criteria. Throughout this work, the central theme remains the improvement of automatic detection and analysis of sand foundry defect, which is a critical factor in modern foundry quality control.
1. Introduction
The foundry industry has always faced challenges in ensuring the quality of cast products. Among various casting defects, shrinkage cavities and porosity are particularly common and can significantly compromise the mechanical integrity of components. In the automotive sector, where aluminum alloy wheels are mass-produced, the need for reliable and efficient inspection methods is paramount. X-ray non-destructive testing (NDT) has become an indispensable tool for detecting internal defects without damaging the product. However, traditional manual inspection of X-ray images is subjective, time-consuming, and prone to human error. This has motivated the development of intelligent X-ray inspection systems that leverage computer vision and image processing techniques to automate defect detection and classification.
The term sand foundry defect encompasses a wide range of imperfections that can occur during the casting process, including gas porosity, shrinkage cavities, inclusions, and cracks. In this thesis, I specifically concentrate on shrinkage cavities (缩孔) and micro-porosity (缩松), which are characterized by their distinct appearances in X-ray images. A shrinkage cavity appears as a large, concentrated bright region with irregular boundaries, while porosity manifests as a cluster of small, dispersed bright spots. Quantitative parameters such as the equivalent diameter for shrinkage cavities and the area density for porosity are used to assess their severity. These parameters directly influence the final quality classification of the casting.
Developing robust defect detection algorithms requires extensive testing with a large variety of defect samples. Collecting such samples from production lines is often impractical because it requires stopping the line, storing thousands of high-resolution images, and dealing with the inherent scarcity of certain defect types. Therefore, there is a pressing need for a simulation framework that can generate realistic virtual defects, which can then be inserted into genuine X-ray images to create synthetic test datasets. Such a framework would not only accelerate the development of detection algorithms but also facilitate the optimization of system parameters in a controlled environment. In the following chapters, I will describe the overall architecture of an X-ray intelligent inspection system, propose a convex hull based method for defect diameter measurement, discuss the application of PDE-based diffusion filtering for image denoising, and detail a novel simulation algorithm for shrinkage cavity and porosity defects. The entire research is conducted with the goal of achieving high accuracy and efficiency in the context of sand foundry defect detection.
2. X-ray Intelligent Inspection System Overview
2.1 System Architecture
A typical X-ray intelligent inspection system consists of two major parts: the hardware front-end inside a radiation shielding room and the software back-end in a control room. The front-end includes an X-ray source, an image intensifier, a CCD camera, and programmable logic controllers (PLCs) for precise motion control. The back-end is a high-performance computer that runs image processing algorithms, controls the hardware through the PLC, and provides a user interface for monitoring and data management. The overall configuration is shown in Table 1.
| Subsystem | Components | Function |
|---|---|---|
| X-ray generation | X-ray tube, high voltage generator | Produces penetrating X-rays with adjustable energy |
| Manipulation | Servo motors, U-arm, conveyors | Positions the casting and rotates the X-ray source for multiple viewing angles |
| Image acquisition | Image intensifier, CCD camera, A/D converter | Converts X-ray shadow into digital image |
| Control | PLC, relays, sensors | Executes motion sequences and sends status signals to the host computer |
| Image processing | Host computer, software algorithms | Performs image enhancement, defect detection, classification, and data storage |
The inspection workflow proceeds through five functional zones. The casting is first placed on an entry conveyor and moves to a centering station where two cameras capture top and side views for model recognition. After determining the wheel type, the system retrieves the corresponding inspection parameters (e.g., X-ray voltage, number of imaging positions). The wheel then enters the X-ray examination chamber where a U-arm rotates around it while the casting itself can also rotate. Multiple X-ray images are captured in real time and immediately processed by the defect detection software. If defects are found, the casting is marked with paint and subsequently sorted into a reject bin.
2.2 Image Processing Pipeline
Defect detection in X-ray images is a multi-stage process. In my research, I follow the pipeline shown below:
Image partition → Preprocessing → Defect extraction → Defect analysis → Casting classification
The image partition stage divides the X-ray image into regions of interest based on the expected defect distribution. For example, the wheel spoke, rim, and hub areas exhibit different background textures and defect probabilities. The preprocessing stage enhances contrast and removes noise. A crucial step in preprocessing is diffusion-based filtering which I will discuss in detail in Chapter 4. Defect extraction involves finding closed contours, filling regions, and filtering out false alarms. It uses a combination of the Laplacian of Gaussian (LoG) operator and dual-threshold edge strength analysis. After extraction, each defect is characterized by its area, diameter, and local contrast. Finally, the casting is classified as acceptable or defective by comparing these statistics with standard thresholds derived from industrial quality standards. The overall process directly contributes to the reliable detection of sand foundry defect in a production environment.
3. Diameter Measurement of Irregular Defects
3.1 Problem Definition
One of the key parameters used to quantify the severity of a shrinkage cavity is its diameter. For a perfectly circular defect, the diameter is easily defined. However, real defects are almost never circular. Instead, their shapes are highly irregular, and a single scalar measure must reflect the spatial extent of the defect. I define the defect diameter as the maximum Euclidean distance between any two points on the boundary of the defect. Geometrically, this is the diameter of the smallest circle that can enclose the entire defect, also known as the Feret diameter. Computing this by brute force would require calculating distances between every pair of boundary points. For a boundary with \(n\) points, the number of distance computations is \(\frac{n(n-1)}{2}\), which becomes computationally prohibitive for high-resolution images with many defects.
3.2 Convex Hull Approach
To reduce the computational cost, I exploit a fundamental property of the diameter of a planar set: the two farthest points always lie on the convex hull of the set. Therefore, I first compute the convex hull of the defect’s boundary point set using Graham’s scan algorithm, which runs in \(O(n \log n)\) time. The convex hull is a convex polygon that contains all points of the original set. After obtaining the convex hull vertices, the diameter can be found by checking only pairs of hull vertices. Because the number of hull vertices is usually much smaller than the total number of boundary points, this results in a significant speedup.
Graham’s scan proceeds as follows. Let \(S = \{p_0, p_1, \ldots, p_{n-1}\}\) be the set of boundary points. The algorithm begins by selecting the point with the smallest y-coordinate (and leftmost among ties) as the anchor \(p_0\). The remaining points are sorted by increasing polar angle with respect to \(p_0\). Then, a stack is used to build the convex hull. For each point \(p_i\) in the sorted list, while the last three points (the top of the stack, the second top, and \(p_i\)) make a non-left turn, the top point is popped. The left-turn test uses the cross product determinant:
\[
T(p_0, p_1, p_2) = \begin{vmatrix}
x_0 & y_0 & 1\\
x_1 & y_1 & 1\\
x_2 & y_2 & 1
\end{vmatrix}
\]
If \(T(p_0, p_1, p_2) > 0\), the turn is strictly left; if \(T = 0\), the points are collinear; if \(T < 0\), the turn is right. This test ensures that only points causing a clockwise turn are discarded. The final stack contains the vertices of the convex hull in counterclockwise order.
3.3 Implementation and Results
In the defect detection algorithm, after the image is binarized (defects are set to 255 and background to 0), connected components are labeled. For each labeled component, I extract its boundary pixels and apply the convex hull algorithm. Then, I compute the diameter by evaluating all pairs of hull vertices. This process is illustrated in the flowchart of Table 2.
| Step | Operation | Description |
|---|---|---|
| 1 | Input | Detected defect mask from X-ray image |
| 2 | Labeling | Find connected components and index them |
| 3 | Boundary extraction | Trace the contour of each component |
| 4 | Convex hull | Apply Graham scan to the boundary points |
| 5 | Diameter | Compute maximum distance between hull vertices |
| 6 | Statistics | Find the maximum diameter among all defects |
I tested the algorithm on a typical wheel X-ray image containing 50 defects. The program completed the entire diameter measurement in approximately 0.35 seconds on a Pentium III 1.01 GHz machine with 384 MB RAM, which is sufficiently fast for industrial online inspection. The convex hull method dramatically reduces the number of distance evaluations compared to a brute-force approach, especially for large and irregular defects. This improvement is crucial because defect diameter is an important criterion for classifying sand foundry defect severity.
4. Image Denoising Using Partial Differential Equations
4.1 Noise in X-ray Images
X-ray images are inherently contaminated by noise from various sources, including quantum noise, electronic noise from the CCD sensor, and quantization noise. Although modern image intensifiers have low noise floors, denoising is still necessary to enhance the detectability of small and low-contrast defects. The denoising process must preserve sharp edges and small details, because loss of these features could lead to missed defects. I found that classical linear filtering (e.g., Gaussian or moving average) blurs edges, and median filtering can distort fine structures. Therefore, I turned to anisotropic diffusion methods based on partial differential equations, which offer a better trade-off between noise suppression and feature preservation.
4.2 Nonlinear Diffusion Models
The most basic diffusion equation is the heat equation:
\[
\frac{\partial u}{\partial t} = \Delta u
\]
where \(u(\mathbf{x}, t)\) is the evolving image and \(\Delta\) is the Laplacian. The solution is equivalent to convolving the initial image with a Gaussian kernel of increasing variance, leading to isotropic smoothing that destroys edges. To overcome this, Perona and Malik proposed a nonlinear diffusion model:
\[
\frac{\partial u}{\partial t} = \operatorname{div}\bigl( c(|\nabla u|) \nabla u \bigr)
\]
with a diffusion coefficient \(c(s)\) that is a non-increasing function of the gradient magnitude, for example:
\[
c(s) = \frac{1}{1 + (s/\beta)^2}
\]
where \(\beta\) is a contrast parameter. In regions where \(|\nabla u|\) is large (edges), the diffusion coefficient is small, thus preserving edges. In flat regions, diffusion proceeds normally and suppresses noise. However, the P-M model can be sensitive to noise because noise creates large gradients that are mistaken for edges. A more advanced approach employs a diffusion tensor to control not only the speed but also the direction of diffusion. The tensor \(D\) is a \(2 \times 2\) symmetric positive-semidefinite matrix whose eigenvectors indicate the preferred diffusion directions. I implemented a model based on time-delay regularization and diffusion tensor:
\[
\frac{\partial u}{\partial t} = \operatorname{div}(D \nabla u) – \lambda (u – I)
\]
\[
\frac{\partial v}{\partial t} + \frac{1}{\tau} v = \frac{1}{\tau} u
\]
Here, \(v\) is a time-delayed version of \(u\), and the diffusion tensor \(D\) is constructed from the structure tensor of \(v\), which is a smoothed version of the gradient. The time-delay regularization smooths the gradient estimates, reducing the influence of noise on the diffusion direction. The parameter \(\lambda\) controls the fidelity to the original image \(I\), preventing over-smoothing. This formulation can be written explicitly as:
\[
u(\mathbf{x}, t) = \frac{1}{\tau} \int_{0}^{t} e^{-\frac{s}{\tau}} u(\mathbf{x}, t-s) \, ds
\]
which shows that \(v\) is a temporal average of \(u\). This averaging helps stabilize the diffusion process and allows the algorithm to preserve fine structural details such as the boundaries of a sand foundry defect.
4.3 Experimental Comparison
I compared the proposed PDE method with several standard denoising approaches on both natural images and wheel X-ray images. The results are summarized in Table 3. The experiments show that the PDE method achieves a better balance between noise reduction and edge preservation. For instance, on a typical wheel X-ray image, the PDE-based method (20 iterations) produced a clearer defect boundary than median filtering or NL-means, while BM3D, though producing a smooth image, destroyed the low-contrast defects that are crucial for detection. In addition, the computational time of PDE (about 620 ms for 20 iterations) is acceptable for online use, whereas BM3D took 35 seconds, making it impractical for real-time inspection.
| Method | Execution Time (s) | Edge Preservation | Defect Visibility |
|---|---|---|---|
| Median filter | 0.08 | Moderate | Partially lost |
| PDE (20 iter) | 0.62 | High | Preserved |
| PDE (50 iter) | 1.10 | High | Too smooth |
| NL-Means | 1.38 | High | Defects nearly invisible |
| BM3D | 35.0 | High | Defects invisible |
Figure 1 below illustrates a typical X-ray image of a casting defect that my simulation framework aims to reproduce.
In the context of sand foundry defect, the denoised image allows the subsequent LoG-based edge detector to produce closed contours around the defect regions. Without adequate denoising, the edge map may contain many spurious closed contours, which lead to false positives. The PDE-based diffusion filter, with proper stopping criteria, minimizes this risk and thus improves the overall accuracy of the intelligent inspection system.
5. Simulation of Shrinkage Cavity and Porosity Defects
5.1 Motivation and Requirements
The evaluation of a defect detection algorithm requires a diverse set of test images containing various defects: different sizes, shapes, contrasts, and locations. Acquiring such a dataset from real production is often expensive and time-consuming. For this reason, I developed a method to simulate X-ray images of sand foundry defect, specifically focusing on shrinkage cavities (缩孔) and porosity (缩松). The simulation should satisfy three main requirements:
- Realism: The simulated defects must resemble real defects in shape and gray-level transitions.
- Randomness: Each simulated defect should be different even when generated at the same location on the same image, mimicking the variability of real defects.
- Accuracy: The size (diameter for cavities, density for porosity) must match the user-specified value within a tolerance (e.g., 10% relative error).
5.2 Gray-Level Statistics of Wheel X-Ray Images
Before designing the simulation algorithm, I analyzed the gray-level behavior of wheel X-ray images. In a small region (e.g., a 100×100 pixel block), the gray values typically vary only slightly. The mean is often around 135, with a standard deviation of just a few gray levels. This means that the background around a potential defect is relatively flat, which is advantageous for inserting synthetic defects without creating obvious seams. In large areas, the gray levels may exhibit periodic patterns due to the wheel’s geometry, such as alternating bright and dark stripes on the rim. The simulation algorithm must preserve these global textures while inserting localized defects.
5.3 Overview of the Generation Model
The overall simulation procedure is depicted in Table 4. It consists of four main steps: template creation, defect shape generation, size adjustment, and gray-level assignment. The algorithm works directly in the spatial domain using the original X-ray image as the background. It reads an image \(f\) and a user-specified location \(A\) where the defect is to be generated. Let \(T\) denote a binary mask (template) of size \(78 \times 46\) pixels (or another fixed size) that defines the maximum area in which the defect can appear. The intersection of the template with the background image gives a set of “base pixels” \(f’\).
| Step | Action | Mathematical Representation |
|---|---|---|
| 1 | Template creation | Design binary masks \(M_1, M_2, M_3\) for shrinkage cavity or one mask for porosity |
| 2 | Shape generation | Select pixels from \(f’\) based on threshold criteria and position functions |
| 3 | Size adjustment | Apply nearest-neighbor interpolation or morphological erosion to fit the target size/density |
| 4 | Gray-level assignment | Increase gray values of defect pixels by 5–10 levels relative to the local mean |
5.4 Shrinkage Cavity Simulation
5.4.1 Nested Templates
To generate a realistic shrinkage cavity, I used three nested binary templates \(M_1 \subset M_2 \subset M_3\), where \(M_1\) is the innermost bright core and \(M_3\) is the outer boundary. The region \(M_1\) will receive the highest gray-level increment, \(M_2\) a moderate increment, and \(M_3\) a smaller increment, creating a bright center that fades toward the edge.
5.4.2 Shape Generation
Let \(e_i\) be the mean gray value of the base pixel image \(f_i’\) corresponding to template \(i\). For a chosen parameter \(x_i \in (0,1)\), I define a threshold as:
\[
\theta_i(m,n) = e_i (1 + x_i) – \lambda \sqrt{m^2 + n^2}
\]
where \((m,n)\) are pixel coordinates relative to the template center. The term \(\lambda \sqrt{m^2 + n^2}\) simulates the radial growth of the cavity. The binary shape \(S_i\) is then obtained by:
\[
S_i(m,n) =
\begin{cases}
1, & f_i'(m,n) \ge \theta_i(m,n) \\
0, & \text{otherwise}
\end{cases}
\]
Experience shows that \(x_1=0.6\), \(x_2=0.65\), and \(x_3=0.68\), with \(\lambda = 0.5\), give good results. The union of the three shapes produces the shrinkage cavity that already has a rough, irregular boundary.
5.4.3 Size Adjustment
The resulting cavity may not have exactly the required diameter \(D_1\). To adjust the size, I use nearest-neighbor interpolation. Let \(S\) be the binary shape and \(D_2\) the actual diameter. The scale factor \(s = D_1 / D_2\) is computed, and the image is resized by \(s\) using nearest-neighbor interpolation. Nearest-neighbor interpolation is preferred over bilinear or bicubic because it preserves the rough, pixelated boundary that is characteristic of shrinkage cavities in X-ray images. The final binary mask after interpolation is denoted as \(S’\).
5.4.4 Gray-Level Assignment
The gray-level increment is applied layer by layer. Let \(a_3 \in [5,10]\) be the increment for the outermost region \(M_3\), \(a_2 \in [3,5]\) for \(M_2\), and \(a_1 \in [3,5]\) for \(M_1\). The total increment for a pixel in \(M_1\) becomes \(a_1+a_2+a_3\). In practice, I choose \(a_3 = 5\), \(a_2 = 4\), and \(a_1 = 4\), resulting in a center brightness 13 gray levels above the local background, which matches the typical contrast of real shrinkage cavities.
After applying the increments, a diffusion process is applied only to the defect region to blend it into the background and achieve a natural transition. The diffusion is a simple linear isotropic process run for a few iterations, which smooths the sharp edges without removing the overall shape.
5.5 Porosity (Shrinkage Porosity) Simulation
5.5.1 Dispersion of Base Pixels
Porosity consists of many small pores dispersed over a larger region. Unlike a single cavity, the gray levels within each pore are almost uniform. I use only the largest template \(M_3\) as the bounding mask. The base pixel image \(f’\) is first transformed by a sinusoidal modulation:
\[
f”(m,n) = f'(m,n) \times (1 + \sin(\sqrt{m^2 + n^2}))
\]
This operation changes the gray values of the base pixels while preserving their spatial distribution. The sine function is chosen because it provides an even distribution of positive and negative perturbations, which helps extract a scattered set of pixels. Then, a threshold \(v\) is applied:
\[
P(m,n) =
\begin{cases}
1, & |f”(m,n) – e_s| < v \\
0, & \text{otherwise}
\end{cases}
\]
Here \(e_s\) is the mean of \(f’\). The threshold \(v\) is typically between 0.5 and 3, and it determines the initial number of seed pixels of the pores. The resulting binary image \(P\) contains very small blobs that will be expanded via morphological dilation.
5.5.2 Morphological Dilation
To make the small seed pixels resemble actual pores, I apply a binary dilation with a disk-shaped structuring element of radius \(R\). The disk is chosen because pores in sand casting are often approximately circular when viewed in an X-ray image. A suitable radius is 4 to 6 pixels; I used \(R=5\). Figure 2 in the original work (omitted here) shows the before and after effect. The dilation combines nearby seeds into larger pores while retaining the overall dispersed character.
5.5.3 Density Control Using Erosion
The density of porosity is defined as the ratio of the total defect pixels to the area of the bounding region. After dilation, the density is often too high compared to the target density \(M_1\). I developed a method called “erosion with shape preservation” to reduce the density without destroying the pore shapes. The steps are:
- Compute the current density \(M_0\) and the required reduction ratio \(r = (M_0 – M_1)/M_0\).
- Label each connected component \(\Omega_u\) and count its pixels \(A_u\). The number of pixels to be removed from region \(u\) is \(T_u = r A_u\).
- Apply a cross-shaped structuring element to erode the region \(\Omega_u\) iteratively. After each erosion, count the total removed pixels \(Z\). Stop when \(Z \ge T_u\).
- Let \(t_u = Z – T_u\) be the number of excess removed pixels. Randomly select \(t_u\) pixels from the boundary of the last eroded layer and add them back to the eroded region.
- Repeat for all connected components. The result is a final porosity mask that preserves the pore shapes while meeting the density requirement.
5.5.4 Gray-Level Adjustment for Porosity
For each pixel in the final porosity mask, I simply add a constant gray-level shift in the range \([5,10]\) to the corresponding pixel in the original image. Because the individual pores are small, this uniform increment is sufficient to produce the desired contrast. No additional diffusion is necessary, as the pores are naturally separated by the background.
5.6 Experimental Results
I conducted extensive experiments using the MATLAB platform. For shrinkage cavities, I set \(\lambda=0.5\), \(x_1=0.6\), \(x_2=0.65\), \(x_3=0.68\). For porosity, I set \(v=1.5\) and \(R=5\). I generated 200 shrinkage cavities and 200 porosity defects on different X-ray images at random locations. The requested diameter \(D_1\) for every shrinkage cavity was 55 pixels, and the requested density \(M_1\) for every porosity defect was 0.0027. The measured diameter \(D_2\) and density \(M_2\) are shown in Table 5 for five representative examples.
| Example | \(D_2\) (pixels) | \(M_2\) | Relative error of \(D\) (%) | Relative error of \(M\) (%) |
|---|---|---|---|---|
| 1 | 54.5260 | 0.0272 | 0.86 | 1.85 |
| 2 | 54.2033 | 0.0276 | 1.45 | 2.22 |
| 3 | 53.9073 | 0.0275 | 1.99 | 2.96 |
| 4 | 54.3783 | 0.0275 | 1.13 | 1.11 |
| 5 | 53.8015 | 0.0277 | 2.18 | 3.70 |
The relative errors remain well below 5%, demonstrating the accuracy of the generation algorithm. To further test the robustness, I varied the requested diameter \(D_1\) over 74 integer values from 7 to 80 pixels. The maximum error was 9%, but the majority of cases were under 3%. Similarly, for the density request ranging from 0.0009 to 0.09, the maximum error was 7%. These results satisfy the requirement of a maximum 10% error and confirm the reliability of the simulation framework for creating realistic sand foundry defect images.
Subjective evaluation by foundry experts confirmed that the generated defects appear natural and closely resemble real shrinkage and porosity defects in X-ray images. The random variations in shape and gray-level transitions are sufficient to challenge the detection algorithm, while the quantitative parameters remain controllable.
Finally, I validated the simulated defects by running them through the same defect detection algorithm used for real defects. The detection algorithm successfully identified both the real and the simulated defects, which is a strong indication that the synthetic images are comparable to real ones from the algorithm’s perspective. This makes the simulation tool highly practical for quickly testing and optimizing detection algorithms without the logistical burden of collecting a large number of real defective samples.
6. Conclusions and Future Work
In this thesis, I have presented a comprehensive approach to the simulation of shrinkage cavity and porosity defects in X-ray images of castings, specifically in the context of sand foundry defect detection. The main contributions of my work are summarized below.
First, I proposed a convex hull algorithm for computing the diameter of irregularly shaped defects. This method significantly reduces the computation time compared to a brute-force distance calculation, while maintaining high accuracy. The diameter is a critical parameter for defect severity classification, and the convex hull method enables real-time analysis in industrial inspection systems.
Second, I investigated the application of partial differential equation-based diffusion filtering for image denoising. By combining time-delay regularization and a diffusion tensor, the algorithm effectively removes noise while preserving the sharp edges and fine details of defects. This pre-processing step is essential for the reliable detection of low-contrast defects such as small porosity.
Third, I developed a novel simulation framework for shrinkage cavity and porosity defects. The algorithm generates defects that satisfy three key requirements: realism, randomness, and accuracy. The shrinkage cavity model uses nested templates and a position-dependent threshold to create irregular shapes with a bright center and fading boundaries. The porosity model uses sinusoidal gray-level modulation and morphological operations to produce a dispersed distribution of small pores. Both models allow the user to specify the target diameter or density, and the actual measurements match the requested values within a small error tolerance. The simulated defects are visually indistinguishable from real defects when inspected by experienced professionals, and they are correctly detected by the existing defect detection algorithm.
Future research directions could include extending the simulation to other types of defects such as gas pores, inclusions, and cracks. Additionally, the current method relies on 2D image features; integrating a 3D model of the casting and the defect would allow more accurate simulation of perspective and attenuation effects. Another interesting direction is to use generative adversarial networks (GANs) to learn the distribution of real defects and then generate even more realistic examples. Nevertheless, the methods proposed in this thesis already provide a solid foundation for the development and evaluation of automatic inspection systems in the foundry industry.
In conclusion, the intelligent X-ray inspection system combined with a reliable defect simulation tool can greatly improve the efficiency and accuracy of sand foundry defect detection. The ability to generate synthetic test images on demand accelerates the development and validation of detection algorithms, ultimately leading to higher quality cast products and safer automotive components.