スカイロジック List of Products and Services

In the manufacturing industry, quality control and quality assurance are essential tasks for maintaining corporate trust. Among the various quality control tasks, "visual inspection" occupies a significant portion. By conducting visual inspections at each stage of the process, the yield of downstream processes can be improved, and the outflow of defects can be prevented in advance. However, this can also increase manufacturing costs and, in some cases, the visual inspection may become a bottleneck, leading to a decrease in production capacity. Here, we will discuss the points to consider for automating visual inspections, the differences between traditional image processing and the latest AI image processing, and trends in visual inspection overseas. ■ What is visual inspection? ■ Main inspection items of visual inspection ■ Automated visual inspection as an alternative to manual inspection ■ Advantages and disadvantages of manual inspection ■ Advantages and disadvantages of automated visual inspection ■ Steps for implementing automated visual inspection ■ Will AI (deep learning) visual inspection replace traditional rule-based image processing? ■ Situations where procedural image processing is used in visual inspection ■ Situations where AI is used in visual inspection ■ Points to consider when automating visual inspection ■ Trends in automated visual inspection overseas ■ Summary ■ Visual inspection systems that can be implemented at low cost *Please see the related links.

I would like to write about various types and characteristics of lighting, as well as their main uses. Please choose the optimal lighting according to your needs. ■ Bar lighting ■ Ring lighting ■ Low-angle ring lighting ■ Backlight (transmitted light) ■ Backlight + polarizing filter ■ Coaxial illumination ■ Dome lighting *For more details, please refer to the related links.

When inspecting with images, the quality of the light that enters the camera's light sensor significantly affects whether the defects you want to detect can be found or not. 1: Influence of Reflected Light One of the factors that changes significantly depending on the angle of illumination is reflected light. For example, as shown in the example below, even when photographing the same object, the appearance can change completely depending on the angle of the lighting. (Figure 1: Left: When illuminated from directly above, Right: When illuminated from an angle) The sample above has a moisture-proof coating applied to the surface of the substrate, giving it an overall glossy appearance. When photographing such an object, if light is directed straight on, the light source may reflect directly, resulting in an undesirable image. In such cases, the lighting is angled. (Figure 2: Left: When illuminated from directly above, Right: When illuminated from an angle) The blue arrows in the above diagram indicate the reflection of the light source. When light is directed from directly above, the reflected light enters the camera directly, but when illuminated from an angle, the reflected light escapes to the opposite side, allowing you to avoid direct reflected light.

Lenses that allow for adjustments in aperture and focus have a value called "focal length," which relates to the angle of view (shooting range). When expressing the specifications of a lens, it is common to represent the maximum aperture value and focal length together, such as f1.4/12mm, which indicates that "focal length" is an important specification. Here, we will introduce how to choose a focal length. *For more details, please see the related links.*

The aperture acts like the iris (brown eye) in the eye, adjusting the amount of light that enters the camera's image sensor, while also affecting the depth of field and the camera's exposure time. This time, I would like to examine depth of field and exposure time. ■Depth of Field ■Exposure Time ■How to Increase Depth of Field While Reducing Exposure Time

As a representative of inexpensive cameras, there are webcams that can be connected to a computer via USB. Their prices start from around 1,000 yen, making them quite affordable, but it's good to understand the reasons why industrial cameras are generally used for image processing to avoid potential issues. Of course, the cheaper option increases cost-effectiveness, but in my experience, there are "3" customers using webcams compared to "7" customers using industrial cameras. Conversely, this means that 30% of customers can operate effectively with a webcam.

"High pixels mean high performance" — Indeed, as the pixel count increases, it becomes possible to capture details more clearly. However, it does not simply mean that larger is better, as high-pixel cameras also have drawbacks. For example, using a camera with unnecessarily high pixel counts can lead to longer image processing times. Therefore, it is necessary to consider whether "it is really impossible to make a judgment without using a high-pixel camera?" when selecting the pixel count of a camera. The example below shows the results of an experiment on how much difference in clarity there is when enlarging a part of a PC motherboard from 300,000 pixels to 14 million pixels. (Refer to the figure) In this way, high pixel performance is demonstrated when trying to see fine details within a wide field of view. However, as mentioned earlier, not only does image processing take longer, but as will be introduced in "High pixels or low pixels (2)," it is also necessary to increase the exposure time. The consideration of whether "it is really impossible to make a judgment without using a high-pixel camera?" is one of the important factors in selecting a camera.

In "(1) High Pixels or Low Pixels," we discussed the advantages of high-pixel cameras in capturing details clearly. Here, I would like to write about the advantages of low-pixel cameras. The advantage of low-pixel cameras can be summed up in one phrase: "processing is faster." There are two reasons for the faster processing. 1. There are fewer pixels to handle as image data. 2. The exposure time can be shorter, resulting in less time needed for shooting. I believe there is no room for doubt regarding point 1. So, why can the exposure time be shorter in point 2? Let me explain. The size of the camera sensor (the square light-receiving part shown in the initial photo) varies, but in industrial cameras, sizes around 5mm square to 7mm square are common (1/3 inch to 1/2 inch). However, as mentioned in "High Pixels or Low Pixels (1)," the variation in pixel counts ranges from about 300,000 pixels to around 14 million pixels. At this time, what differences exist in the size of the pixels?

As explained in "Whether High Pixel or Low Pixel (2)," industrial cameras come with various sensor sizes. Generally, they range from 1/3 to 1/2 inch, but there are also sizes like 1 inch. A larger sensor size means that, with the same number of pixels, the pixel size increases, resulting in: - A larger light-receiving area per pixel, allowing for shorter exposure times. - Less noise in images when using the same exposure time. - Sharper images with the same lens focusing performance (it might be easier to understand if you think that a larger pixel size makes the focus target larger). These are the advantages (however, larger sensor sizes are more expensive to produce, so the price also increases accordingly). One point to note is that if you want to change the sensor size from 1/2 inch to 1 inch to improve image quality, the angle of view (field of vision) will also change, even if you use the same lens.

The types of camera light sensors can be broadly divided into two categories: "global shutter" and "rolling shutter." The difference between these two lies in whether all the pixels of the sensor are exposed simultaneously (global shutter) or if they are exposed one line at a time from the top down (rolling shutter). How does this difference in mechanism affect the image? A typical example is the distortion shown below. When capturing a moving object, with a global shutter, a spherical object will be photographed as a sphere, but with a rolling shutter, because exposure occurs line by line from the top, the object may have moved during the exposure time, resulting in distortion. Thus, the differences between the two become apparent when photographing moving subjects. One might wonder, "Wouldn't it be better to use only global shutters?" However, in terms of manufacturing costs, rolling shutters are more advantageous (lower cost). (Figure 1: Left: Global Shutter, Right: Rolling Shutter)

One of the common concerns when selecting a camera is the interface. In recent years, most manufacturers of industrial cameras have released both GigE and USB3.0 cameras. This is because each of the two interfaces has its own advantages and disadvantages. In this article, I would like to discuss how to choose the appropriate interface based on your application. ■ Price Comparison The differences between the two are clearly defined in the specifications, so I will extract some of that information. By looking at the table below, you may clarify which camera you should choose for your current application. (Note: Manufacturers offer cables up to 5m or 8m, but there may be cases where operation becomes unstable.) ■ Power Supply for GigE Cameras When connecting to a PC via GigE, you cannot power the camera directly from the PC even if you connect the PC and camera with a LAN cable. Therefore, you need to either power the camera directly or use a separate PoE injector or PoE hub between the PC and the camera to supply power to the camera. (Note: PoE hub capable of powering 4 cameras (BS-GU2005P))

This time, I would like to organize the types of AI. A new introduction page for DeepSky has also been created, so please take a look. Now, the term "AI" that is currently in the spotlight refers to "deep learning," which is the type of AI that can distinguish between pedestrians and traffic signs, and even defeat human professionals in chess and Go. In the world of image processing, there are also machine learning-based image processing and what is called procedural image processing. So, what are the differences between them? I think it will be easier to understand the meanings of each by looking at the diagram below. AI (Artificial Intelligence), as the name suggests, means artificial intelligence and has been the most widely used concept since around 1950, meaning "something that replaces human intelligence." In the world of image processing, "procedural" (rule-based) image processing, which processes images taken by a camera according to a set procedure to determine OK/NG, also falls under this category of AI. Our EasyInspector uses this type of AI across a wide range, including color judgment and dimensional angle inspection.

When talking with customers, I sometimes feel that they believe AI (deep learning) automatically determines what is OK or NG. This is a subtle nuance, and it's not something that needs to be corrected in the course of actual conversation, but today I want to clear up that confusion by writing this article. Now, regarding the confusion mentioned above, the actual difference is that it is humans who decide what is considered OK and what is NG in the software. This leads to the question, "So what does AI do then...?" To put it simply, what our AI does is "find what it has been taught," and fundamentally, that’s all there is to it. This function is generally referred to as "object detection." In "object detection," it literally detects "objects" within an image.

Currently, I am learning a lot about images, and I realize the importance of "taking good pictures." Cameras come equipped with various functions, and it seems that if you can set them correctly, you can capture good images. One of those settings is "white balance." As someone who was a novice in image processing, I had heard of it but wondered what it actually does. After looking into it, I found it interesting and potentially useful, so I would like to share it. *For more details, please see the related links.*

Recently, we have been receiving more inquiries from customers using AI software (such as EasyInspector2 and DeepSky) asking questions like, "Why does this happen?", "Isn't it supposed to be like this?", and "What is going on inside?". In such cases, it seems that the person in charge is unsure about how to explain things. Indeed, when trying to provide an intuitive and easy-to-understand explanation, they often end up relying on analogies, or if they attempt to explain in detail, they find that specialized books provide more thorough information. I have also been unable to find suitable explanatory materials for customers who want to take a deeper dive into the principles of AI image processing.

We often receive requests to use high-resolution cameras to improve detection capability. In the case of rule-based image processing, using high-resolution images tends to improve resolution and enhance detection capability, but this is not always the case with deep learning. Below is a brief explanation of the perspective on resolution in deep learning, albeit in a rough manner. Let's consider images like (1) to (3) in Figure 1. (1) Total area 10×10, area of the gray rectangle 4 (2) Total area 20×20, area of the gray rectangle 16 (3) Total area 10×10, area of the gray rectangle 16 The area of the gray rectangle in (2) is four times larger than in (1), but when looking at the ratio of the gray rectangle to the total area, (1) is 4/100 and (2) is 16/400, both representing only 4%. In terms of "ease of detecting the gray rectangle" in deep learning, (1) and (2) are almost the same.

This time, I would like to verify the "differences in learning and inspection speed using GPUs" as stated in the title. I will compare the differences in learning and inspection times with two types of GPUs and 16GB and 8GB of RAM. ■Conditions ■Verification Configuration ■Results *For more details, please see the related link (blog).

This is something that our sales engineers explain almost every time they talk to customers, but despite being a crucial factor that greatly influences AI performance, many of the discussions found online are too general to be practically useful. We thought that if we could provide a more practical, specific, and realistic explanation, it would surely be helpful to everyone, which is the background of this blog. ■ No matter how good the AI is, if it is taught incorrectly, it will produce incorrect results - About teacher images - About annotations - Convenient features in annotations - Blogs and videos that are helpful *For more details, please see the related links (blog).

The manufacturer we consulted this time is a company that has been inquiring with us for some time. They sent us sample images. Our company also conducts usage explanations and demonstrations via web conference. First, it would be smoother to enter the verification process after asking about the client's issues and operational status during the web conference. ■ Inspection Settings and Results We conducted learning and processing on the two types of images provided, one with spectral lighting and the other with general lighting. For both images, we specified abnormal areas and trained the system to identify them. In the spectral lighting image, we were able to detect some potentially defective areas, and I feel there is a possibility to improve detection accuracy by increasing the training data. In the general lighting image, it honestly seems difficult to make judgments. It appears challenging to detect anything other than large, grainy defects like those in the image on the right. The left image shows the detection of defects in the spectral lighting image. The numbers in the detection image represent the AI's confidence level in percentage, which we refer to as the number of recognition points.

One common question from companies considering the introduction of image recognition and inspection using deep learning is, "How many images do we need to prepare for inspection?" The conclusion is that it cannot be definitively stated. This is because the amount of data required varies significantly depending on the complexity of the object's appearance (color, shape, angle, etc.) and the changing features. In this article, we have organized the basic verification process as follows: 1. Capture images of the product to be inspected and collect approximately a few hundred images (for example, around 200). 2. Select half of those images and provide "annotations" for the object. 3. Use the annotated data as training data for the AI and validate it with the remaining images (testing for recognition). 4. If there are misrecognitions or missed recognitions, increase the number of images or review the annotations and retry. 5. Repeat this "data augmentation → training → validation → readjustment" process until satisfactory accuracy is achieved. *For more details, please refer to the related link (blog).

When implementing an image recognition and inspection system using AI (deep learning), we often encounter the question, "How much training data is needed?" To cut to the chase, the answer is "there is no fixed number," but the key to successful implementation lies in "using good data appropriately." This article explains the significance and utilization of training data as follows: - The roles of training data, validation data, and evaluation data - Designing to prevent data bias - The required amount of data depends on "variability factors" - Label accuracy and annotation design - Ongoing maintenance after operation *For more details, please see the related link (blog).

We will introduce common challenges heard from customers and ways to solve them. Challenge 1) It is difficult to analyze why a judgment was incorrect. Challenge 2) I want to quantitatively determine whether the completed model is good or bad. Challenge 3) When learning, I want to visually know which items are likely to be confused. Challenge 4) I don't know when to stop learning. When does overfitting begin? Challenge 5) Defective products are not being collected for new products. *For more details, please see the related link (blog).

Recently, there has been an abundance of inexpensive network cameras available, which has allowed our EasyMonitoring2 system (a system that collects images from network cameras for tasks such as meter reading) to be implemented at a lower cost. However, this has also brought about certain issues. Specifically: - Most cameras have become of a type that does not allow lens replacement. - The predominant lenses are those designed to cover a wide area, which often results in significant image distortion (similar to fisheye lenses). Due to these trends, it has become challenging to capture images of meters from a distance using inexpensive cameras. Since the EasyMonitoring2 system connects to a large number of cameras, there is a demand for compact and affordable options. Therefore, we decided to test a method that allows for capturing images of meters even with low-cost cameras. *For more details, please refer to the related link (blog).

This time, we used a PTZ camera with the aim of capturing multiple locations with a single camera. A PTZ camera is one that has the following mechanisms: Pan: to move the camera horizontally Tilt: to move the camera vertically Zoom: to zoom in (optically changing the angle of view). (Images are a general representation of PTZ cameras.) Let's quickly check the field of view. *For more details, please refer to the related link (blog).