Visual inspection is indispensable for quality control in a wide range of industries, including manufacturing and infrastructure management. While visual inspection was traditionally the mainstream, AI-based visual inspection has also come into use in recent years, enabling higher-precision automated inspections.
In this article, we explain the methods of visual inspection, the advantages and disadvantages of AI systems, and points to consider. By reading until the end, you will understand the systems used in visual inspection and the key considerations.
|
【Table of Contents】 |
Visual inspection is an inspection method to maintain quality by checking the surface of products and parts for defects. Scratches, dirt, and shape abnormalities are detected using visual inspection or automated systems.
The purpose of visual inspection is to discover defects and failures in the appearance of products and parts at an early stage to maintain quality and increase customer satisfaction. Visual inspection is essential to avoid the risk of defective products circulating in the market and damaging a company's brand value and trust.
Furthermore, it is possible to discover deficiencies in the manufacturing process or mechanical abnormalities early through visual inspection, which can also improve the efficiency of the production line.
Therefore, the development and implementation of a visual inspection system is an important decision-making process directly linked to a company's productivity and quality control.
Inspection Items in Visual Inspection
The main inspection items in visual inspection are as follows. Although specific items vary based on the manufactured goods and company standards, inspecting these items allows problem-free products to be sent to market.
In the visual inspection of shape, the following points are mainly checked to ensure the product maintains its designed form.
Shape is an important inspection item because it affects not only the product's appearance but also its functionality. Especially in products with complex shapes, minute shape deviations determine quality.
In the visual inspection of structure, checks are made to ensure that product parts are correctly assembled and that there are no issues with strength or durability. It is also necessary to inspect for misalignments in assembly and combination.
Strict inspection is required because structural defects carry the risk of leading to failure or damage during transportation or use of the product.
In dimensional inspection, measurements are taken to see if the product is manufactured with the dimensions according to the design drawings. Dimensional errors can interfere with the product's operation, necessitating precise measurement that meets standards.
There are mainly two types of visual inspection methods:
Each method is explained below.
Manual visual inspection is a visual inspection method where workers directly check products using their eyes and sense of touch to discover defects or abnormalities.
Inspection by the human eye excels at finding parts that are difficult for machines to capture, such as fine scratches, subtle color changes, and the adhesion of foreign matter. It can also respond flexibly to different inspection criteria for each product.
However, manual inspection has limits. Because the precision of manual inspection depends on worker fatigue and experience, variations in inspection accuracy and consistency may occur.
Therefore, since it becomes difficult to maintain quality through manual inspection alone for products requiring mass production or high precision, it is often used in combination with automated inspection.
Automated inspection is a method of inspecting automatically using machinery or AI technology. Cameras and sensors are used to obtain images of products or equipment, and defects or abnormalities are detected based on set standards.
Automated inspection is efficient, especially in mass production, because it can inspect a large volume of products in a short time. Additionally, inspection accuracy and consistency are high, and inspection precision does not vary due to human fatigue or subjectivity.
Because it is possible to develop automated inspection systems that can detect even minute defects or dimensional deviations that would be missed by the human eye with high precision, it is suitable for products requiring high quality or parts with complex shapes.
However, the content of automated inspections is limited to pre-set inspection rules, and AI basically cannot inspect content outside of the range it has learned. Furthermore, responding to high-mix low-volume production may require system adjustments or the implementation of multiple systems.
In visual inspection, the appropriate system differs depending on the purpose and precision. There are several types of visual inspection systems as follows:
Each is explained below. It is not uncommon for these devices to be used together within a single line, leveraging their unique advantages.
Desktop-type visual inspection systems are a method using compact, space-saving inspection units. An operator manually sets the product (inspection target) to perform the inspection.
These are utilized in small-scale production lines or sites seeking simplification of inspection. They are easy to implement because they can be used simply by being placed on a table or workbench.
Because of their simple structure and relatively easy operation, it is possible to perform inspections with a certain degree of precision while keeping initial costs low.
Semi-automated visual inspection systems combine human labor with machine automation. When an operator sets a product, the device automatically controls rotation and lighting. Then, the operator performs the inspection visually. In some cases, part of the inspection or the recording of results is performed on the system side.
These can respond more flexibly than complete automation while increasing efficiency compared to fully manual visual inspection. They are suitable for cases where product types are diverse and flexibility in manual adjustment or inspection is required.
Automated-type visual inspection systems are fully mechanized systems that perform the entire sequence of steps from product inspection to judgment automatically. By automating, it is possible to inspect large volumes of products quickly and accurately, making them suitable for large-scale production lines operating 24 hours a day or precise visual inspections.
In automated types, by combining cameras, sensors, image processing technology, and AI, it is possible to reliably detect minute defects and dimensional deviations that are often overlooked by visual inspection.
Especially in systems utilizing AI, complex pattern recognition is possible, and it is not uncommon to achieve far higher precision than manual inspection. Additionally, because human error and variations in judgment by inspectors can be eliminated, consistency in inspection quality can be ensured.
Offline
Offline visual inspection is a method of inspecting products by installing an inspection device in a location separate from the production line. Products are temporarily removed from the line, and visual inspection is performed with a dedicated inspection instrument. This is a system type that stands in contrast to the inline visual inspection described later.
Since manual inspection is used in many cases and inspection is performed in a dedicated inspection space, the advantage is that detailed inspections can be carried out thoroughly.
Inline visual inspection is a method of implementing visual inspection within the production line. Inspection devices are designed to fit the production line, allowing inspection to be implemented without stopping the flow of manufacturing. Achieving 100% inspection is also relatively easy.
Inline inspection can maintain a smooth flow of the production line, allowing defective products to be detected in real-time.
AI-based inspection systems are a method of achieving high precision and efficiency by automating visual inspection through the utilization of AI image recognition technology.
By utilizing deep learning technology to learn vast amounts of image data, it is possible to possess high inspection capabilities that recognize minute defects and regularities. Therefore, defective products and abnormalities that are difficult to detect with traditional inspection systems can be accurately detected.
Furthermore, AI can learn from image data obtained through visual inspection to improve inspection precision. The longer the system is operated, the more it is expected to significantly improve productivity and enhance the uniformity of quality.
However, precision is greatly determined by how much sufficient learning data regarding defective products and abnormalities can be secured during AI training. Additionally, the selection of appropriate hardware (cameras, lighting, etc.) is also important.
3D inspection systems are technology for inspecting products and parts three-dimensionally. Product shape data is obtained using 3D cameras and sensors.
Since traditional 2D inspection only provides planar information, there was a possibility that depth or complex shapes could not be inspected. 3D inspection captures the overall picture of the product, enabling detailed and simultaneous detection down to shape deviations and surface irregularities. Especially when combined with AI, complex pattern recognition can also be performed.
Mobile inspection devices are battery-powered, portable visual inspection systems equipped with IoT sensors and cameras.
Because they can be easily moved and used inside and outside the factory, they are effective when line changes occur or when inspections need to be performed per product. They are particularly valued in small-scale production lines or sites that need to frequently change inspection locations.
Additionally, they recently come equipped with small high-precision sensors and infrared cameras, making it possible to perform high-quality visual inspection regardless of location. Furthermore, some devices utilize AI technology, allowing more advanced analysis to be performed on the spot.
4. Advantages and Disadvantages of AI-based Visual Inspection
In recent years, the utilization of AI for visual inspection has increased.
However, to implement AI, considerations must be made regarding technical and cost aspects. Here, we explain the advantages and disadvantages of utilizing AI for visual inspection.
Advantages of AI-based visual inspection include the following:
The advantage of AI-based visual inspection lies in the fact that it significantly improves the efficiency and precision of inspection. Because it learns from vast image data and can detect fine defects and abnormalities with higher precision than traditional methods, it can reliably detect spots that are easily missed by the human eye.
Furthermore, the speed of inspection is improved, making it possible to inspect large volumes of products in a short time.
In addition, AI-based visual inspection combines the strictness of traditional rule-based image processing technology with the flexibility of manual visual inspection by humans.
Traditional visual inspection using image sensors had low strictness in drawing the line between OK and NG, and was poor at inspecting areas known as "gray zones." On the other hand, AI is capable of flexible judgment close to human manual inspection even in situations called gray zones.
Moreover, AI inspection can operate continuously for 24 hours and contributes to reducing errors caused by worker fatigue or sensory variations, thus improving the consistency of quality control. As a result, it also leads to the reduction of production costs.
AI-based visual inspection has the following disadvantages:
Building an AI-based system that can withstand long-term practical use requires high-performance hardware and software, as well as a large volume of learning data. Preparing these requires a large investment at the time of implementation.
Long-term operational costs and the return on investment period (ROI) should be closely considered in advance.
Furthermore, because AI performance depends not only on the quantity but also the quality of learning data, there is a risk that inspection precision will drop if appropriate data is not prepared. Additionally, preparation of data including data collection and annotation requires specialized knowledge, and the time it takes until the start of operation can also be a disadvantage.
Also, AI is not "omnipotent." Since AI makes judgments based on learned data, there is a possibility it cannot accurately detect completely new types of defects not included in the learning data. Consideration is also required for the possibility that additional learning data or system readjustment may be needed to adapt to new defect patterns or changes in the manufacturing environment.
When implementing AI in visual inspection, you must pay attention to the following points:
Each point is explained below.
For an AI system to function in visual inspection, patterns of defects or abnormalities in the products or parts to be inspected must be defined, and it must have learned from appropriate data corresponding to those patterns. Therefore, it is indispensable to confirm in advance whether the inspection items match the AI's inspection performance.
For example, if the judgment of the implemented AI deviates from the company's quality standards or industry specifications, inappropriate pass/fail judgments will be made. Furthermore, if an AI model with a slow processing speed relative to the production line speed is implemented, inspection will become a bottleneck and production efficiency will drop.
If implemented without confirming suitability with the inspection content, expected quality control cannot be achieved, and conversely, costs and man-hours may increase.
In image recognition used for visual inspection, annotation is the work of labeling image data. Labels such as "Normal," "Defective," or the defect's "Type," "Degree," or "Position" are assigned to the image data used for learning.
If annotation is inaccurate or performed based on ambiguous standards, the risk of the AI making incorrect judgments and lowering inspection precision increases. Therefore, ensuring the quality of annotation in image recognition is indispensable for improving AI system precision.
Automated annotation technology is progressing in the field of image recognition as well. Even so, the current reality is that it falls far short of the flexibility and precision of manual annotation using human eyes and hands.
To perform and maintain high-precision annotation, it is necessary to document definitions of defects and abnormalities and classification standards in detail and share them among annotators. Additionally, mutual checks by multiple annotators are performed to exclude individual subjectivity and errors.
In cases where manufacturing processes or types of defects diversify, it is also important to maintain annotation quality through regular data updates and verification.
For methods and procedures of image annotation, please also see: "How to annotate Images: A thorough explanation of the methods, procedures, points to note, and whether to outsource or do it in-house!"
By investigating cases of other companies in the same industry, you can grasp in advance unique challenges and difficult points in products or manufacturing processes similar to your own. This makes it easier to predict problems that could occur during the introduction of AI for visual inspection in your company and to formulate countermeasures.
When investigating other companies' cases, the collection of failure cases becomes as important as or more important than success cases. By analyzing success cases, you can learn effective implementation methods and operational best practices.
And by studying other companies' failure cases, you can reduce the risk of repeating the same mistakes. By analyzing the causes of failure and reflecting them in your company's implementation plan, you can formulate a more solid plan.
In the development and implementation of an AI-based visual inspection system, you must choose between in-house or outsourcing.
In the case of in-house, because the development to operation of the AI visual inspection system is performed consistently within the company, it is possible to build an optimized system. You can operate a system specialized only for your company's manufacturing process.
However, it is necessary to have specialized knowledge of AI technology and annotation within the company, and there is a disadvantage that initial costs and human resource development take time and effort.
On the other hand, in the case of outsourcing, because you can utilize the knowledge and technology of specialized providers, the advantage is that you can implement a high-precision system in a short time. Especially for companies that do not have AI-related know-how, they can build systems that could not be realized on their own.
Detail a list of what can and cannot be done within the company, including the aforementioned annotation work part. Depending on the company's situation, a hybrid approach—starting with outsourcing in the initial stage and gradually moving to in-house production—can also be an effective option.
Utilizing AI for visual inspection enables efficient and accurate inspections. For companies that want to improve the quality of visual inspection or implement an automated inspection system, why not base it on AI?
Annotation is important for making the AI models used in visual inspection high-performance. If you have concerns about the implementation or quality of annotation, consider consulting an annotation specialist company.