スカイロジック List of Products and Services

In recent years, very strict quality control has been required, and even slight scratches on products are treated as defective. Therefore, it has become important to detect scratches more reliably. Our company proposes a low-cost system that sounds a buzzer when defective products are detected on the conveyor, stops the conveyor with a red light on the patrol light, and resumes operation once the defective products are removed. This is a very practical and easy-to-understand system, so please make use of it. [Materials and Components (Excerpt)] ■ Solid State Relay: AQA211VL ■ DC24V Adapter: Akizuki Electronics / GF18-US24075T ■ Outlet: Panasonic / WH2163KWP ■ Wooden Board: 300 x 200 x 12 ■ Buzzer: Akizuki Electronics / PB10-Z338R *For more details, please refer to the PDF document or feel free to contact us.

We will introduce a proposal to detect forgotten parts flowing on the conveyor, smooth out the subsequent processes, and prevent the shipment of defective products. By applying a system where a buzzer sounds and stops the conveyor when an NG (non-conforming) item is detected, we can simultaneously inspect two types of forgotten nuts and screws. We will guide you on how to register settings using the specific image processing software "DeepSky," which stops the conveyor when an NG is detected and resumes operation once the NG is removed. For more details, please check the related links below. 【Tools and Equipment】 ■ PC: Mouse Computer/G-Tune, Core i9, 16GB RAM, RTX 2070 SUPER ■ Image Processing Software: Sky Logic/DeepSky DS100K ■ IO Unit: Sky Logic/EI-ITIO-T01 ■ 1.3 Megapixel Camera: Daheng/MER-133-54u3c ■ 12mm Lens: M1214-MP2 ■ Camera Stand (Aluminum Frame) *For more details, please refer to the PDF document or feel free to contact us.

We would like to introduce a case where color mistakes in electronic harnesses were detected using the image processing software "DeepSky." By trying multiple settings and annotations (the process of surrounding the inspection area to teach the software what to detect) to distinguish color mistakes, it is possible to achieve a high level of detection stability. We have provided information on the differences between annotations that could not be judged and those that could be judged through the related links below. Please take a look. 【Materials and Components (Excerpt)】 ■ Solid State Relay: AQA211VL ■ DC24V Adapter: Akizuki Electronics / GF18-US24075T ■ Outlet: Panasonic / WH2163KWP ■ Wooden Board: 300 x 200 x 12 ■ Buzzer: Akizuki Electronics / PB10-Z338R *For more details, please refer to the PDF materials or feel free to contact us.

We would like to introduce Inaba Electric Industry's "Choco Stop Watcher mini" within the context of dashcam-like cameras. If an abnormality occurs that can be detected by the device's sensors, it can simply be input as a contact signal to the "Choco Stop Watcher mini" to record video. However, when the abnormality is due to "dirt" or "angle misalignment," it is often difficult for the sensors to detect the issue. Therefore, this time we tried to "detect abnormalities using image processing." When an object is placed at an incorrect angle, it is detected as an abnormality through image processing, and recorded by the "Choco Stop Watcher mini." For more details, please check the related links below. 【Set Examples】 ■ Image Processing Unit + Choco Stop Watcher mini ■ Multiple Choco Stop Watcher minis connected to a single PC *For more details, please refer to the PDF document or feel free to contact us.

I believe there are many instances where people visually check for insects or hair on food flowing on a conveyor. Here, I would like to introduce a system that incorporates AI vision to more reliably detect foreign objects. By controlling the ON/OFF of a 100V power supply with a relay, the conveyor itself is turned OFF when a foreign object is detected, stopping the conveyor and simultaneously sounding a buzzer to alert about the foreign object detection. This system uses DeepSky and Intelligent I/O, with a relay connected to the NG output of the Intelligent I/O. For more details, please check the related links below. 【Tools and Equipment】 ■ PC: Mouse Computer/G-Tune, Core i9, 16GB RAM, RTX 2070 SUPER ■ Image Processing Software: Skylogic/DeepSky DS100K ■ IO Unit: Skylogic/EI-ITIO-T01 ■ 1.3 Megapixel Camera: Daheng/MER-133-54u3c ■ 8mm Lens: M0814-MP2 ■ Camera Stand (Aluminum Frame) *For more details, please refer to the PDF document or feel free to contact us.

We will introduce a system that detects broken tablets and stops the conveyor. It was found that using a mechanical relay to perform ON/OFF at 100V generates significant electromagnetic noise, so we switched to a solid-state relay for testing. We added a patrol light to the wiring for "ON/OFF of the 100V power supply from DeepSky," which alerts with a red light and buzzer when a broken tablet (defective product) flows through, stopping the conveyor. When good products continue to flow, a green light is illuminated, and the conveyor keeps moving. For more details, please check the related links below. 【Tools and Equipment】 ■ PC: Mouse Computer/G-Tune, Core i9, 16GB RAM, RTX 2070 SUPER ■ Image Processing Software: Sky Logic/DeepSky DS100K ■ IO Unit: Sky Logic/EI-ITIO-T01 ■ 1.3 Megapixel Camera: Daheng/MER-133-54u3c ■ 25mm Lens: M2514-MP2 ■ Camera Stand (Aluminum Frame) *For more details, please refer to the PDF document or feel free to contact us.

This time, we are inspecting the top surface and the entire circumference of the cap using only one camera (without using multiple cameras). We will introduce a case where the friction force is utilized to rotate the cap. When the AI image inspection software DeepSky detects a defect, an NG signal is output from the Intelligent I/O, and by connecting an electromagnetic valve to the NG output, the valve opens when a defect is detected. At this time, air from the air compressor flows to the pusher, causing the pusher to extend forward and automatically discharge the defective product. For more details, please check the related links below. 【Tools and Equipment (Excerpt)】 ■PC: Mouse Computer/G-Tune, Core i9, 16GB RAM, RTX 2070 SUPER ■Image Processing Software: Sky Logic/DeepSky DS100K ■IO Unit: Sky Logic/EI-ITIO-T01 ■1.3 Megapixel Camera: Daheng/MER-133-54u3c ■12mm Lens: M1214-MP2 *For more details, please refer to the PDF document or feel free to contact us.

This time, I would like to focus on the production of a mechanism called "string conveyor." Unlike a regular conveyor, a string conveyor transports products along a "line" rather than a "surface." Why do we go to the trouble of supporting products with a line? The advantages are: - It allows for inspection from both the top and bottom since it can be photographed from below. - By placing a backlight underneath, it enables photography using transmitted light. - By varying the speeds of the two strings, it can rotate the products. And so on. Production time: 1.5 days Material cost: 30,000 to 40,000 yen [How to make it] This time, there will be no cutting or gluing. The basic approach is to assemble the various parts that have been arranged using a screwdriver and hex wrench. *For more details, please refer to the related links or feel free to contact us.*

We received a question from a customer using the image inspection software EasyInspector, asking, "Can inverter fluorescent lights not be dimmed?" We inform customers using EasyInspector that "if there are no issues with inspections using indoor fluorescent lights or inverter fluorescent lights, there is no need to use expensive lighting." Many LED lights have separate lighting and power units, and since the power unit can be dimmed with a knob, I thought such questions were natural. In fact, there are various methods for adjusting brightness: 1) Dimming 2) Aperture of the lens (in the case of industrial cameras) 3) Camera exposure time and gain settings In other words, in principle, if any of the above three can be adjusted, brightness can be controlled. So, what is the best method to use for adjustment? (Figure 1) This is a photograph of part of a bottle cap. The focus is set on the bottom of the cap. Both images have the same brightness, but you can see that the left image has the focus extending to the edge of the cap compared to the right. This is referred to as "having a deep depth of field."

Function: OK/NG Judgment by Area This is a function that judges not only "how many items were detected in the image" but also "where they were detected." Depending on the application, it may not be possible to make a correct judgment based solely on the total count on the screen. For example, in the case where there are two capacitors positioned alternately, as shown below. In this case, if the orientations of the capacitors are reversed, they should be deemed unacceptable; however, as for the total count, there would be one with the polarity mark on top and one with it on the bottom, resulting in both being considered acceptable. (Figure 1) In such cases, judgments are made by dividing into areas. As shown in the dotted box below, predetermined areas are set up, and the count judgment is performed for each area. (Figure 2) Within each area: "If one upward-facing capacitor is detected, it is OK." "If even one downward-facing capacitor is detected, it is NG." Settings like these are made. (Figure 3) This allows for correct NG judgment even for defective items that have the same count but are positioned alternately. (Figure 4)

DeepSky is an image inspection software that uses so-called AI (Deep Learning). By training it on the parts you want to detect, the software adjusts its own setting parameters and learns to recognize them. Here are three features that specifically differentiate it from traditional methods. *We will use the inspection of washer scratches as an example.* ▼ No need for positioning of the inspection target (Figure 1: No scratches, OK) (Figure 2: Scratches, NG) The fact that positioning is unnecessary means that inspections can be performed with the same settings even if the number of washers on the screen varies. This is effective for inspection targets that are difficult to secure. ▼ Can detect even with changes in brightness (Figure 3) It can detect scratches even when the lighting is this dim. This is effective in cases where it is difficult to suppress reflections from metal parts or when dealing with products that come in multiple colors.

▼Feature: Auto Annotation This is a function that automatically executes annotations. It was previously a beta feature, but it has become an official feature starting from Ver. 2.2.0.0. Many of you may know that annotations are very important in object recognition. However, when there are many target objects in an image, the amount of work required for annotation increases, and as the workload increases, mistakes such as forgetting to annotate or making errors inevitably rise. If annotations are forgotten, DeepSky adjusts parameters to determine that "this is something that should not be found (even though it looks similar) for the same object," which can significantly lower the recognition rate of the object, leading to various issues. (See Figure 1: Annotation Forgetting) This is where the auto annotation feature comes into play. By using auto annotation, with just a click of a button, it automatically suggests annotations like "Based on the previous training data, you would want to annotate here, right?" (See Figures 2, 3, and 4)

In recent years, with the miniaturization and precision of machinery and equipment, smaller components are often used, and when performing image inspection of such small-sized parts, high-resolution cameras are increasingly utilized for imaging. However, when using high-resolution cameras, a high-spec computer is also required to connect the cameras. Therefore, in this report, we would like to present the operational verification and results of connecting two high-resolution (14 million pixels) cameras to a single computer and starting them simultaneously. As a result of the verification, we were able to display the screens of both 14 million pixel cameras on a single computer screen. (Figure 1) Additionally, the image below shows the image captured by one of the two connected cameras. It is possible to detect parts that are less than 2mm (approximately 0.8×1.6mm) within a field of view of about 100mm square. (Figure 2 shows an enlarged image of the part outlined in red in the overall image on the left.)

DeepSky has the capability to communicate the coordinates of detected objects to a higher-level system via TCP/IP, enabling various applications by tracking the coordinates of objects within that system. Here, we introduce software that inspects objects flowing on a conveyor and counts them when they cross the center of the camera. The diagram below (Figure 2) illustrates the configuration of the conveyor, camera, and software. The camera's images are processed by DeepSky to detect the type and coordinates of the objects, and this information is passed to the upper counting software. The counting software tracks the objects and increments the count when they cross the center of the image. This is an image of products moving on a belt conveyor. The upper camera captures images while checking for defective items and counting the products. It can also interact with PLCs to reduce speed or trigger a buzzer when approaching a specified count. The counting software tracks the coordinates of the objects to count them. In this way, DeepSky can work in conjunction with higher-level software, allowing for various uses. Custom software, like the counting software, can also be developed by our company.

During the verification and after the implementation, we received many valuable questions and reports from users regarding the settings of DeepSky, so we would like to introduce them. Q. How do you determine overfitting? Q. What does a convergence of 0.1 or lower mean? Does it become harder to converge as the number of labels increases? Q. What is the difference between continuing training from a certain point and resetting and retraining? *For more details, please refer to the related links (blog).

This time, I will introduce information regarding annotations. Annotations are one of the important settings related to detection accuracy, so I hope you find this useful. Q. Should labels be grouped together or divided into finer categories? When there are multiple detection targets (such as scratches, dirt, and dents), there are two patterns: registering all annotations under the same label "defect" or dividing the labels by shape as "scratches, dirt, dents." DeepSky tends to show improved detection results with fewer labels, so it is generally better to register them under the same label. However, if it is necessary to know which defect has been detected, it is essential to separate the labels by shape. In this case, if a scratch is mistakenly registered as dirt, it can lead to inconsistencies during training, resulting in poor learning outcomes. Therefore, annotations must be carried out carefully to avoid mistakes and oversights. (Figure 1)

We sometimes receive inquiries from customers regarding the specifications of PCs when connecting multiple devices. This time, we would like to introduce the behavior when connecting multiple cameras using the AI function of EasyInspector2 (hereinafter referred to as EI2). We hope this will serve as a reference when selecting a PC. Additionally, we hope it will also provide guidance on the expected increase in inspection time when connecting multiple devices. We used a multi-controller (hereinafter referred to as EIMC) to start and inspect six EI2 units. ■PC ■Hardware Configuration ■Software Configuration ■Results *For more details, please refer to the related links (blog).

In the mechanism for turning the 100V power ON/OFF created in 'Discovering Attached Insects and Stopping the Conveyor' (https://skylogiq.co.jp/DIY_HowTo/291), a large mechanical relay was used. By turning the contacts with a potential difference of 100V ON/OFF using the relay's electromagnetic coil, unexpectedly large electromagnetic noise was generated. Electromagnetic noise can cause communication failures and other issues in interfaces such as USB. This time, I used a solid-state relay (SSR) to create a power ON/OFF mechanism that minimizes electromagnetic noise. At the beginning of the video, a ring fluorescent light is connected to the 100V power, and the process is as follows: DeepSky makes an OK/NG judgment → OK/NG output to Intelligent I/O → the buzzer and solid-state relay connected to Intelligent I/O turn ON/OFF → the ring fluorescent light turns ON/OFF. If a conveyor belt is used instead of the ring fluorescent light, the conveyor will move when OK and stop (with the buzzer sounding) when NG, enabling automatic operation.

We received an inquiry from a customer who manufactures automotive parts and industrial machinery regarding the settings of DeepSky. They asked, "I thought the system was searching the entire screen based on the characteristics of the annotated area I set, but do you also remember any tendencies about where it tends to appear in the field of view?" 【Inspection Settings and Results】 We arrange squares of the same color and size and annotate only the upper part for training (see the image in the top left). In the image on the top right, we can find only the upper part. In the lower left image, we can also find only the upper part. However, in the right image, we could not find anything from the lower part. If the system can only find data that closely matches the teacher due to overfitting, the situation changes a bit. However, the position of the items we want to detect, such as defects, greatly affects inspection accuracy. It is necessary to use teacher images with various positions, orientations, and angles. As of 2021, DeepSky has been equipped with a teacher image augmentation feature. We can augment images of defects that rarely appear by flipping or rotating them in all directions and using brightness or Gaussian noise.

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).

It can be combined with rule-based functions. When trying to measure the area or dimensions of a target object within a complex background, the background interfered, making it difficult to detect edges as desired. However, by extracting the target object and binarizing it using AI segmentation, inspections have become much easier. *For more details, please see the related link (blog).

Recently, we released a new product, the "AI Sheet Counter ai-Numbers." ai-Numbers is a new counting machine that uses AI to handle sheet materials where counting has been difficult due to variations in cross-sections caused by the state of cutting blades and significant differences in thickness and appearance based on the material. Today, we will introduce the differences between the counting machine "ai-Numbers" and "EasyNumbers." In manufacturing, "counting errors" directly lead to reduced yield and complaints. Image processing-based counting machines can resolve this issue at once, but the optimal solution varies depending on whether the edges of the counting targets are complex and difficult to understand or if the focus is on counting visually clear edges as quickly as possible. We will compare our counting machines, ai-Numbers and EasyNumbers, in terms of technology, application areas, and performance, and summarize the key points for selection. *For more details, please refer to the related link (blog).

In conventional image processing, it is possible to inspect the degree of terminal insertion, but there are challenges such as limited connector shapes and the need for positioning. Therefore, we decided to use AI for the inspection. By training the AI to recognize uninserted, partially inserted, and fully inserted states, it becomes possible to identify them. While the AI object recognition function is sufficient for "detection," the real issue is how to make the judgment. AI object recognition establishes inspection by having humans create settings that search for and recognize defects, such as scratches or dents, from the overall workpiece—if even one defect is found, it is deemed a failure (NG). In other words, object recognition alone can identify an object, but it cannot make a judgment that it is acceptable just because it appears in a certain alignment. This is where the OCR function comes into play. *For more details, please refer to 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).