Automated Defect Detection through Flaw Grading in Non-Destructive Testing Digital X-ray Radiography

Abstract

Process automation utilizes specialized technology and equipment to automate and enhance production processes, leading to higher manufacturing efficiency, higher productivity, and cost savings. The aluminum die casting industry has significantly gained from the implementation of process automation solutions in manufacturing, serving safety-critical sectors such as automotive and aerospace industries. However, this method of component fabrication is very susceptible to generating manufacturing flaws, hence necessitating adequate non-destructive testing (NDT) to ascertain the fitness for use of such components. Machine learning has taken the center stage in recent years as a tool for developing automated solutions for detecting and classifying flaws in digital X-ray radiography. These machine learning-based solutions have increasingly been developed and deployed for component inspection, to keep pace with the high production throughput in manufacturing industries. This work focuses on the development of a defect grading algorithm that assesses detected flaws to ascertain if they constitute a defect that could render a component unfit for use. Guided by ASTM 2973-15; Standard Digital Reference Images for Inspection of Aluminum and Magnesium Die Castings, a grading pipeline utilizing K-D (k-dimensional) trees was developed to effectively structure detected flaws, enabling the system to make decisions based on acceptable grading terms. This solution is dynamic in terms of its conformity to different grading criteria and offers the possibility to achieve automated decision making (Accept/Reject) in digital X-ray radiography applications.

1. Introduction

Process automation utilizes specialized technology and equipment to automate and enhance production processes. It combines hardware, software, and information technology to manage and control manufacturing operations, leading to higher manufacturing efficiency, greater productivity, and increased cost savings [1]. The aluminum die casting industry has significantly gained from the implementation of process automation solutions in manufacturing, leading to remarkably high production throughputs of aluminum die casting components [2], serving industries such as the automotive and aerospace industries. This manufacturing technique involves injecting molten aluminum alloy into a mold, cooling the mold, and extracting the component from the mold. This production method is very susceptible to generating manufacturing flaws, due to the processes involved, and requires adequate monitoring [3]. Nonetheless, aluminum die casting (ADC) has several advantages over other manufacturing processes [4], including the lighter weight of the product, competitive cost for high-volume requirements, higher production throughput, higher process automation possibilities, and better product consistency.

Considering how unsafe it could be to allow defects in manufactured aluminum die casting to go undetected, it is essential to perform adequate NDT inspection of the manufactured parts required [5]. NDT plays a significant role in a variety of industries as it offers a variety of techniques that allow for the detection of flaws in a component, without causing damage or compromising the component’s functionality. Common NDT techniques include radiography, Ultrasonic Testing, visual inspection, Infrared Thermography, Liquid Penetrant Testing, Eddy Current Testing, etc. [6]. According to research, radiography stood out as the most preferred NDT method out of the different NDT methods used to evaluate the quality of aluminum die casting parts [7]. The century-old radiography procedure uses X-rays or gamma rays with intensities high enough to penetrate materials, allowing for a thorough examination of the entire volume of a component under test [8]. X-rays are invisible to natural human vision, thereby necessitating the use of sensors such as digital detector arrays (DDAs), photon-counting detectors, X-ray films, etc., to reveal the latent information contained in the X-ray signal during image acquisition.

1.1. Digital X-ray Radiography

Digital radiography (DR) is increasingly replacing the conventional use of film radiography. In this mode of imaging, X-ray photons captured by the sensitive units (pixels) of the detectors are converted into numerical values. Furthermore, for easier visual appeal and human comprehension, the 2D matrix of pixel values are converted to gray level distribution, leading to the formation of digital radiographs. Digital radiography images are useful for computational processes [9,10,11,12]. Hence, different post-acquisition adjustments of image properties (e.g., perceptible contrast, sharpness, brightness) as well as other image filtration processes are possible with digital radiography images. To ensure the adequate performance of DR systems and to ensure full integration of digital images for use in NDT, relevant standards have been established. Incorporating NDT standards into radiographic testing is crucial for ensuring accurate and reliable inspections across various industries. NDT standards provide essential guidelines on radiographic techniques, exposure parameters, image quality, and result interpretation. These standards are widely applied in industries like aerospace and automotive, where detecting flaws such as cracks or inclusions is critical for safety and quality assurance. Understanding and adhering to these standards enhances the consistency and reliability of radiographic inspections, making them an integral part of the NDT process. Bodies such as The American Society of Mechanical Engineers (ASME), American Society for Testing and Materials (ASTM), and International Organization for Standardization (ISO), offer regulatory oversight for standardization of practices within the NDT industry. The ideal case for NDT practices in the industry requires that NDT inspection is carried out by qualified NDT inspectors that have satisfied the requirements of the operational NDT Standard on qualification of NDT inspectors, e.g., ISO 9712. Such trained inspectors acquire and interpret radiographic images following relevant operational NDT Standards [13]. After an image quality is considered acceptable, inspectors zero in on relevant detected indications to evaluate and make decisions about whether components are to be accepted or rejected (see Figure 1). Despite the regulations in place to prevent potential errors in the NDT practice, it would be fallacious to rule out the possibility of NDT inspectors, being human, making errors [14,15]. The interpretation of radiographic images by an NDT inspector may be influenced by factors such as fatigue, inspection experience, and state of mind. Hence, even after implementing a structured set of regulatory guidelines, a study found the effectiveness of visual human interpretation to be approximately 80% [16].

With the increasing computational capacity over the recent years, ideas to automate non-destructive testing (NDT) processes have been conceived and developed, offering a viable pathway to achieving 100% inspection of manufactured parts. The use of Artificial Intelligence solutions to automate the recognition and evaluation of flaws from NDT data has been a widely prevalent approach explored by many researchers in NDT for the identification of flaws in radiographic images of materials [17]. These AI-based solutions are required due to the growing need to automate tasks traditionally performed by human operators [18]. Considering the increasing demand for NDT in the process-automation-powered manufacturing industries and the decreasing number of qualified NDT inspectors to satisfy this growing inspection need, the necessity to use AI in non-destructive testing becomes more apparent [19]. Although computer-vision-based solutions can be beneficial, they must be developed and applied in accordance with relevant operational NDT standards to foster acceptance in safety-critical sectors.

1.2. Aluminum Die Casting

Aluminum die casting components are susceptible to having manufacturing flaws [20], the majority of which could be invisible on the surface of the cast component, but within the volume of the casting [21]. As a result of the existing capability for automating die casting production, industries currently have a significant production throughput, which makes it challenging to test every manufactured component, due to the influence of confounding factors such as time, cost, and the shortage of qualified NDT personnel to undertake inspection tasks. Several industries have adopted statistical approaches to inspect only a given number of manufactured parts, even though such an approach stands inferior to the 100% inspection of all manufactured parts. Therefore, to achieve 100% inspection of all manufactured components, the idea of utilizing computer-based alternatives has become prevalent, due to its recorded successes in other spheres of application. A variety of industries use aluminum die casting (ADC) to manufacture components, offering superior dimensional precision for near net shape and minimizing the need for extensive secondary machining [22]. Due to the intended functions of certain aluminum die casting components in service, their failure could cause significant economic loss for the industries and customers, rendering them prone to litigative repercussions.

2. Literature Review

Automated defect recognition (ADR) in non-destructive testing (NDT) using digital X-ray radiography is becoming an essential component of quality control in aluminum die casting. Recent publications highlight the importance of ADR technologies, emphasizing their significance in the evolving landscape of quality assurance. Our work, which focuses on grading flaws, is positioned strategically. The transition from merely detecting and characterizing flaws to systematically grading them to determine if they meet standards is crucial for advancing the technology and increasing its industrial acceptance. In recent years, advances in ADR algorithms have been seen in the literature, highlighting the possibility of developing solutions that are capable of computationally addressing the NDT inspection needs in the industry, particularly related to aluminum die casting. In a study [23], object detection methods (based on YOLO, RetinaNet, and EfficientDet) were used to detect aluminum casting defects. The research showed high performance results on the testing dataset, with an average precision of 0.90 and an F1 factor of 0.91 achieved by the trained model, demonstrating the model’s ability to assist human operators in identifying defects in aluminum castings. Similarly, another study [24] investigated the use of deep learning for defect identification, utilizing YOLOv3_EfficientNet object detection method that replaces the backbone of YOLOv3_darknet53 with EfficientNet. Their work demonstrated good model performance. Similar success was recorded using computer vision and deep learning to streamline flaw detection [25]. Convolutional Neural Networks (CNNs) play a particularly significant function in this domain. Ref. [26] introduced a CNN model that is highly efficient and designed to detect defects in radiographic images. They demonstrated the importance of continuous refinement of neural network architectures to enhance the precision of detection. Collectively, these studies reveal a substantial trend toward the utilization of sophisticated computational methods and deep learning to enhance the performance of ADR systems.

Nevertheless, several obstacles and constraints continue to exist, despite these developments. Complex defect geometries and varying radiographic conditions are frequently a challenge for traditional algorithms, as discussed by [26]. This is further exacerbated by the extensive training data that are necessary for deep learning models to generalize effectively across various flaw types and casting conditions, a point that was raised in [23]. Furthermore, the authors of [25] observed that the high computational requirements of advanced CNNs present substantial obstacles to real-time industrial applications; hence, they proposed an efficient CNN model for detecting tiny casting flaws using image-level labels. Their work utilized an object-level attention mechanism with bilinear pooling to enhance defect detection ability.

While substantial progress has been achieved in flaw identification and characterization, defect grading remains an understudied but critical field. Accurate grading is critical for determining the severity and impact of detected flaws, which, in turn, determines fitness for use of the casting under testing. The ability to precisely grade flaws enables producers to categorize them according to severity and potential impact on product performance, improving the reliability and safety of aluminum castings used in essential applications such as automotive and aerospace components. Integrating defect grading into existing ADR systems allows for a more complete approach to quality assurance. Manufacturers can obtain a comprehensive understanding of manufacturing processes by combining detection, characterization, and grading, resulting in higher product quality and less waste. This comprehensive approach is especially important given the research gaps found in the current literature. Our work seeks to fill this gap by presenting a novel flaw grading methodology that succeeds the detection and characterization capabilities of existing flaw recognition solutions. This breakthrough is projected to close the gap between spotting errors and understanding their practical ramifications, resulting in a strong framework for quality control in aluminum die casting.

By filling this gap, our article aims to contribute to the evolution of ADR technologies, ensuring that aluminum castings satisfy the highest quality and reliability criteria. This work emphasizes the need for continual innovation in ADR systems and the potential impact of incorporating defect grading into these systems.

2.1. ASTM2973-15 Standard Digital Reference Images for Inspection of Aluminum and Magnesium Die Casting

A reference document ASTM E2973-15 Standard Digital Reference Images for Inspection of Aluminum and Magnesium Die Casting provides a guide for the visual recognition of specific categories of casting discontinuities and their corresponding severity levels. Porosity, Cold Fill, Shrinkage, and Foreign Materials are accounted for, with corresponding digital reference images. Identifying defect occurrences is crucial, as such data can be used for auditing and optimizing the manufacturing process to reduce the likelihood of recurrence during future manufacturing [27].

In this work, the grading of the severity of detected flaws shall be performed in conformance to terms contained in ASTM 2973. Figure 1 depicts the radiographic image evaluation workflow for NDT. It is important to mention that these standards were intended for visual inspection by qualified NDT technicians, who rely on their subjective interpretation of the acquired X-ray images in comparison with the corresponding reference images. Although the reference images help to reduce interpretation errors, the randomness of flaw size, shape, and distribution within the test samples coupled with overlapping features, variations in material thickness, and image projections directly impact image interpretation and therefore make it more challenging to automate the grading process. Numerous research efforts have been dedicated to identifying, characterizing, and qualifying flaws in digital radiography images, like [17,28] which evidenced the ability of their solutions to detect flaws (referred to in their works as defects). However, there has been no presentation of a defined flaw grading process, according to operational NDT Standards (or any other), to qualify the detected flaws as defects. In other similar studies, detected discontinuities are addressed as flaws [29,30,31].

The use of specific terms in non-destructive testing (NDT) is crucial for comprehending and conveying the outcomes of testing procedures. As per ASTM E1316, the recognized terminology for NDT, the definitions of discontinuity, flaw, and defect are clearly outlined and hold key roles in assessing materials and components.

• Discontinuity: A lack of continuity or cohesion, an intentional or unintentional interruption in the physical structure or configuration of a material or component (ASTM E1316). Discontinuities can arise intentionally, like in design elements or during manufacturing, or unintentionally due to material defects or irregularities in the process. For example, during casting, various discontinuities like gas holes, shrinkage, or foreign particles may occur. Understanding these discontinuities is crucial in interpreting their importance, especially in aluminum casting.
• Flaw: An imperfection or discontinuity that may be detectable by non-destructive testing and is not necessarily rejectable (ASTM E1316). This term signifies an imperfection or break that can be detected through NDT but may not require rejection. A flaw indicates an irregularity that could impact material properties but does not automatically mean the material is unsuitable for use. Distinguishing between a flaw and a defect is critical, as not all flaws compromise the component’s performance. For instance, a small, isolated inclusion in an aluminum casting may be visible through digital radiography but might not affect the part’s overall functionality, thus classified as a flaw rather than a defect.
• Defect: One or more flaws whose aggregate size, shape, orientation, location, or properties do not meet specified acceptance criteria and are rejectable (ASTM E1316). A defect is a significant problem that affects the component’s usability or safety. Defects undergo thorough examination as they can lead to failures during use. In aluminum casting, defects like large shrinkage cavities, cracks, or excessive porosity beyond industry standards must be rectified to prevent potential product failures.

Our study focuses on assessing defects in aluminum casting through grading flaws that are categorized in their unique classes. It highlights the necessity of a systematic approach in assessing the integrity of materials. These procedures utilize machine learning, and classical computer vision, as elaborated in the literature, to accurately pinpoint anomalies in digital X-ray images. Based on the definitions considered, it can be firmly asserted that not all imperfections qualify as defects. This distinction is crucial as many components can still be utilized in service despite known acceptable imperfections. Therefore, the motivation to develop a grading algorithm aligned with the ASTM 2973 Standard arises, considering the steps essential to evaluating an image before classifying detected imperfections as acceptable or rejectable (defects).

This work focused on four (4) categories of flaw indications in digital X-ray radiography images of aluminum and magnesium die casting, as contained in ASTM E2973-15 Standard Digital Reference Images for Inspection of Aluminum and Magnesium Die Casting, which provides a guide to enable recognition of specific casting discontinuity types and their corresponding severity levels. Figure 2 shows reference sample images from ASTM2973-15, showing Porosity in aluminum die casting at distinct levels of severity, ranging from the least (1) to the highest (4).

Each circular area in the digital reference image represents a 700 mm² region that represents increasing severity levels ranging from 1 to 4. The example shown in Figure 2 above is for Porosity in aluminum die casting.

2.2. Objective and Research Purpose

• To develop a comprehensive flaw grading algorithm to classify detected flaws into different severity levels, ensuring alignment with reference standards (client-specific or by standardization bodies).
• To enhance the integration of automated defect recognition (ADR) solutions in safety-critical sectors and propose strategies to address existing barriers.
• To implement computer vision techniques that can yield deterministic outcomes, improving the accuracy and interpretability of outcomes.

3. Materials and Methods

The algorithm developed in this work addresses the processes that succeed defect detection, segmentation, and characterization, leading up to an Accept/Reject decision. Hence, the input to our grading algorithm is ideally an output segmentation mask of the flaws that are detected and classified by an algorithm trained to perform such tasks. Our approach considers developing the flaw grading algorithm as an independent pipeline from the defect detection and characterization pipelines, to make their operations mutually exclusive. Hence, our proposed solution’s robustness is not limited to the performance of the previous detection and characterization stages. Typically, our proposed solution is designed to take as input the instance segmentation which identifies and segments each distinct object within an image, providing the class labels. Hence, when a radiographic image is passed through an already existing automated flaw detection and characterization pipeline, such an algorithm outputs the detected and characterized flaws as images. This output is used in our proposed solution to grade the flaws to ascertain if they constitute a defect. In our test cases, four classes of flaws were considered, in accordance with ASTM 2973. Several test cases were curated from ground truth annotations of real flaws from aluminum die casting parts.

The flaws were extracted from multiple images to form a repository of flaws of varied sizes and morphologies. The extracted flaws were used to populate our input images, randomly assigning 4 distinct color codes to differentiate the 4 different flaw classes considered in the ASTM E2973-15 Standard. The distinctiveness of the colors used offers the proposed grading algorithm the ability to adequately detect each flaw and ensure their subsequent separation of the entire flaws into respective classes. Details of the color codes that were utilized are provided in

Table 1. Distinct representation of flaw classes in input image.

Flaw Category	Color	RGB Values
Porosity	Blue	[223, 34, 26]
Shrinkage	Yellow	[5, 251, 251]
Cold Fill	Purple	[223, 26, 116]
Foreign Material	Orange	[46, 141, 254]

. With this approach of synthetically generating the test cases, a large variety of flaw distributions were realized and used to effectively test and validate our flaw grading pipeline. A schematic description of the methodical approach used to generate the input images is described in Figure 3.

3.1. Grading of Flaws

This work involves the inspection of aluminum die casting components. The input being considered (see Figure 4) shows a random distribution of flaws. The grading reference considered in this work has four (4) distinct severity levels: Grade 1, Grade 2, Grade 3, and Grade 4 in ascending severity order. The evaluation terms considered in this work are contained in ASTM E2973 Standard, presented as follows:

• A 700 mm² area is selected for evaluation. All the flaws within a selected area shall be classified, counted, and sized (area and length).
• When the severity level of discontinuities in the production image being evaluated is equal to or less than the severity level in the specified reference image, that part of the casting represented by the production image is acceptable. If the production image shows discontinuities of greater severity than the reference image, that part of the casting is rejected.
• When two or more categories of discontinuity are present in the same production image, the predominating discontinuities, if unacceptable, shall govern without regard to the other categories of discontinuity, and the casting is rejected.
• When two or more categories of discontinuity are present to an extent equal to the maximum permissible level as shown in the applicable standards for each category, then that part of the casting shall be judged unacceptable.
• There are no limiting criteria for a single size of discontinuity, a maximum number of discontinuities per unit area evaluated, specific dimensional spacing, and/or alignment criterion between individual discontinuities or any other undefined discontinuity patterns.
• There is no limit as to the extent of acceptable discontinuities in a casting, provided that no unit evaluation area throughout the casting contains discontinuities that exceed the severity of discontinuities in the applicable reference image.

To utilize the ASTM references are images, there is a need to convert these reference images to a digital metric that can be easily deployed for the purpose of comparison with the measurements conducted on the input images. Our solution, based on relevant terms contained in ASTM 2973, is versatile enough to accept custom reference values to be agreed to by the client. These values are presented in

Table 2. Custom reference used in this study showing different severity grades.

Flaw Type	Grade	Flaw Area (mm2)	Flaw Quantity	Area [x] (mm2)
Porosity	1	700	1	0.12 < x ≤ 1.0
	2		5	1.0 < x ≤ 3.0
	3		4	3.0 < x ≤10.0
	4		3	x > 10.0
Shrinkage	1	700	1	0.12 < x ≤ 1.0
	2		4	1.0 < x ≤ 3.0
	3		4	3.0 < x ≤ 10.0
	4		2	x > 10.0
Cold_fill	1	700	1	0.12 < x ≤ 1.2
	2		5	1.2 < x ≤ 2.0
	3		4	2.0 < x ≤ 6.0
	4		3	x > 6.0
Foreign body	4	700	1	x > 0.12

and can easily be updated depending on the task being considered. The exact terms used for determining flaw grading (e.g., flaw area and quantity within an inspection area of 700 mm²) were decided and used as reference values. Although our solution uses skimage.measure.regionprops library to acquire more than 30 properties of each blob, in this use case, only the area was utilized. Other properties could be used depending on the task and client’s recommendation for reference values.

Table 2 above describes a typical example of a custom reference that could be given by a client in the industry. Four classes are considered, and there is a possibility of having 4 different severity levels in 3 of the flaw classes (Porosity, Shrinkage, and Cold Fill), while the detected presence of a foreign body is aways having the highest possible severity level of 4 (according to the Standard). To attain a specific severity level, there are conditions that must be met—the area of the detected features, and the quantity of such flaws within a specified area of 700 mm² on the part under investigation.

To have a test dataset that will be sufficiently representative of real-world scenarios, test images were generated with random placement of the 4 flaw classes (Porosity, Shrinkage, Cold Fill, and Foreign Body) that significantly vary in morphology and size within a predefined region of the image. To execute the grading of flaws in the input images, several key steps are particularly important, including the algorithm’s identification of flaws according to different classes (Porosity, Shrinkage, Cold Fill, and Foreign Body), followed by the creation of a class-based binary mask retaining the exact spatial positions of the detected flaws. Therefore, for each given input image, binary images were generated, with each one containing only a specific class of defect that was separated by the color coding in the input image. Furthermore, these binary images were individually processed as input for a class-based assessment of the properties of flaws belonging to that specific flaw class, followed by the subsequent grading steps. Hence, in the case described in this work where four flaw classes were considered, the grading process is applied for a maximum of four times if all flaws are detected in the input images. It is important to mention that when applied to already existing flaw detection and characterization algorithms, a synchronization must be realized between the output representation of such algorithms and the input search of our proposed algorithm. In Figure 5, a schematic representation of the methodology utilized in this work is presented.

3.2. Evaluation Efficiency

Our study introduces a novel application of the K-D tree for efficient data organization and accurate flaw assessment within input images. The use of the K-D tree algorithm encompasses a series of critical phases essential for the successful implementation of this methodology. The K-D tree (short for k-dimensional tree) is a binary search tree used for organizing points in a k-dimensional space. It operates by recursively partitioning the space into nested half-spaces, allowing for efficient range searches and nearest-neighbor queries. The construction of a K-D tree begins by choosing a dimension (typically alternating among the k dimensions) and sorting the data points along that dimension. The median point is selected as the root, and the data are split into two halves—those that lie to the left of the median and those to the right. This process is repeated recursively for each half, building the tree by splitting along successive dimensions. Once constructed, the K-D tree enables efficient querying by quickly narrowing down the search space [32,33]. It is important to acknowledge that K-D trees are most effective in low-dimensional spaces, and their performance can degrade as dimensionality increases, a phenomenon known as the “curse of dimensionality”. However, our application remains firmly within a 2D space, where the K-D tree performs optimally. Even if the methodology were to be generalized for higher-dimensional feature spaces, the nature of the blob features—primarily tied to their spatial coordinates—would not lead to significant dimensional growth that could compromise the K-D tree’s efficiency.

Initially, the spatial distribution of defects is established through preprocessing the processed class-based binary image to identify and extract attributes related to flaw features. Subsequently, a K-D tree is constructed using the centroids of these features, enabling the efficient detection and retrieval of flaw features based on their spatial coordinates. This approach ensures a comprehensive analysis of flaw severity through the implementation of effective radius queries to capture all defects within a specified distance from a given point. Moreover, the management of regions of interest (ROIs) is enhanced by employing a sliding window technique to effectively control and coordinate ROIs. The incorporation of the K-D tree facilitates rapid defect detection within an evaluation ROI, which, in turn, expedites the analysis and evaluation processes. The proposed approach reduces computational complexity, making it particularly advantageous for handling large images with numerous defects. The integration of ROI management enables real-time analysis of individual regions. In essence, the K-D tree assumes a critical role in this algorithm as a foundational data structure that facilitates efficient spatial queries crucial for effective flaw detection and evaluation. To efficiently evaluate the distribution of flaws without expending time on regions with no defect of interest within the input image, this stride function was incorporated to dynamically navigate through the sorted list of flaw features achieved using the K-D tree algorithm. This process is employed to efficiently sort and search for flaw features along different axes (x and y axes in 2D space of the input image). The stride function dynamically navigates through the input image, ensuring that all flaws are inspected and graded for their severity in an efficient manner. The stride function achieves this by calculating the distance needed to move the ROI along a specified axis to reach the next flaw.

The selection of the K-D tree over other data structures is grounded in its superior performance for handling multidimensional spatial queries, particularly in the 2D space of our input images. While alternatives like R-trees [34] offer their advantages, they tend to incur higher computational costs and complexity, especially when managing many small regions. The K-D tree strikes a balance between efficiency and complexity, making it ideal for our application, where the primary need is for rapid and accurate retrieval of flaw features based on their spatial distribution. Furthermore, the implementation of a stride function enhances the efficiency of the K-D tree by dynamically navigating through the input image, ensuring that all flaws are inspected and graded with minimal computational overhead.

4. Results

The output of this grading pipeline is presented as an overlay of activated regions of interest within the processed input image that has a flaw distribution that meets the reference terms utilized in this study. The results are presented according to flaw classes to ensure better visualization and easier human validation. Four severity levels are considered, each assigned a region-of-interest (ROI) color for identification. The color codes for grade 1 through grade 4 are blue, green, orange, and brown. Each flaw class is treated separately, and the grading terms used are as described in Table 2. Grading is achieved per unit area of 700 mm² on the input image as documented in the reference document. When multiple activated ROIs overlap, the combined areas of the ROIs are seen in the results as presented in Figure 6. When activated regions of interest (ROIs) corresponding to different grading terms overlap, the ROI associated with the higher severity level is assigned visualization precedence. The Foreign Body class, if detected, returns as grade 4 (brown ROI) in accordance with the reference terms in Table 2.

5. Discussion

The solution presented in this work addresses notable misunderstandings between flaw detection and defect detection solutions, which are sometimes incorrectly used interchangeably within the field of non-destructive testing (NDT) research. Accounting for this differentiation during image evaluation is important as not every flaw influences the functionality or safety of a component. Hence, mislabeling flaws as defects can lead to unwarranted rejections, and inspection inefficiencies. A flaw or group of flaws constitutes a defect only when it surpasses specific predefined tolerance thresholds or specifications contained in agreements (e.g., Standards), thereby rendering the component unsuitable for its intended purpose.

The solution presented in our work assesses the severity of flaws within a set unit area of interest against industry-approved standards and classifies them as defects only if they fulfill the criteria. This method guarantees that only those flaws that genuinely impact the component’s usability are categorized as defects, thereby streamlining the inspection process. The grading system employs advanced image processing methodologies and a rule-based structure to evaluate various characteristics of identified flaws, such as size, position, and nature. By combining these factors with industry-specific standards, the system can dynamically adapt its grading criteria, rendering it highly adaptable to various inspection scenarios and client requirements.

Our solution is designed to integrate seamlessly with existing flaw detection systems, requiring only minimal modifications. These modifications ensure that the output of detection and characterization from current algorithms is formatted to align with the settings of our grading algorithm, such as assigning unique color codes to each class of flaw present in the image. This approach eliminates the need for a complete overhaul of existing algorithms. Additionally, the methodical design of our solution allows for the customization of the pipeline to meet the specific demands of various industrial sectors, ensuring precise and reliable defect detection. Furthermore, the modular design of our system enables effortless updates and personalization. As standards evolve or new inspection requirements arise, the grading criteria can be easily adjusted. This flexibility ensures that our solution remains relevant and efficient over time, serving as a valuable enhancement to existing flaw detection algorithms.

6. Conclusions

This study re-examines defect detection by emphasizing the critical distinction between flaws and defects. By incorporating a robust grading system into flaw detection algorithms, we offer a detailed, reliable, and adaptable approach to identifying defects from detected flaws in digital radiographic images. Our approach decouples the evaluation process outlined in ASTM 1316, focusing on grading relevant flaws to make informed decisions about the presence or absence of a defect, based on established non-destructive testing (NDT) standards used in the industry. The grading pipeline proposed in this work not only enhances the efficiency and accuracy of the inspection process but also provides a pathway for fully automating decision making and quality control in the radiographic evaluation of manufactured components (Accept/Reject). By seamlessly integrating this method into existing NDT workflows, industries could see a significant reduction in inspection times and costs, while maintaining or even improving the accuracy and reliability of defect grading. This strategy emphasizes the importance of adhering to operational NDT standards in defect grading, ensuring that only a flaw (or group of flaws) with the potential to compromise a component’s fitness for use are classified as defects, based on agreed-upon standards. This solution represents a significant advancement in the field, offering a versatile and adaptable tool for various industries.

Future research could explore the upgrading of the proposed grading algorithm to integrate into 3D applications for volumetric images, such as computed tomography images. Additionally, the incorporation of machine learning-based techniques to further refine the grading criteria could diversify the applicability of the proposed approach, enabling even greater accuracy. These directions could broaden the scope of our proposed approach, solidifying its role in the next generation of automated NDT digital radiography solutions.