As products and components grow more precise, compact, and advanced, they require increasingly high levels of quality. Quality inspection, such as appearance inspection, is therefore becoming more sophisticated throughout the entire product manufacturing process, from raw material manufacturing to product processing and assembly. Toshiba has an extensive track record of real-world deployment of appearance inspection devices using image processing technologies developed through long years of research and development. It has also created AI technologies capable of advanced automated inspection with shorter deployment times by using the non-defect learning method. Toshiba's optical inspection technologies perform one-shot visualization of minute defects invisible to the naked eye. Let's learn about Toshiba's quality inspection automation solution, which combines these devices and technologies.

The Japanese manufacturing industry has worked hard to prevent defective products from being shipped to market. The key to ensuring product quality is quality inspection, such as appearance inspection, and until recently this has primarily consisted of sensory inspection in the form of visual inspection. However, this "visual inspection," performed using the human eye, results in inconsistent inspection results depending on the experience and physical condition of individual inspectors, which creates the problem of defective products being overlooked.

In recent years, products have become precise, compact, and sophisticated. This has led to dramatically higher quality requirements for products and the components and materials that make them up.  In conjunction with this, expectations for the scope of quality inspections have also grown. They are being called on to detect even more minute defects, produce greater product quality uniformity, and rapidly identify the factors that cause defects. In addition to the problems already faced by conventional visual inspections, inspections must now also be able to detect defects on the micron level, invisible to the eye. Visual inspections are reaching their limits.

In response to these rising product quality demands, Toshiba offers a quality inspection automation solution that leverages advanced image processing technologies.

Toshiba's appearance inspection devices form the core of this quality inspection automation solution. They are combined with optical inspection technologies and proprietary Toshiba AI technologies that enhance their capabilities and functions.

In the manufacturing industry, in the past, appearance inspection was normally performed using visual inspection. Inspection results could be inconsistent, and defective products were sometimes overlooked. To solve these problems, the manufacturing industry has been shifting from visual inspection to automated inspection using appearance inspection devices. Through this, companies are seeking not only to improve product quality, but also to better manage quantitative quality data and reduce the burden placed on inspection personnel.

Toshiba has been supplying appearance inspection devices based on our unique image processing technologies for 35 years. Over that time, we've improved device capabilities and functionality as we supported the appearance inspection operations of customers in the manufacturing field. Our devices have a track record of wide-ranging use in material manufacturing fields, used for chemical products such as film and tape, and also with iron, nonferrous metal, woven fabric, nonwoven fabric, paper, electrical materials, LCDs, pharmaceutical packaging materials, and more. Many companies, both inside and outside Japan, have used our appearance inspection devices.

Toshiba appearance inspection devices, with their proven record of use with various materials, use cameras to detect the presence and type of surface defects. They use proprietary inspection algorithms and lighting technologies developed through our years of experience, together with our material-specific operation know-how.

By improving appearance inspection accuracy, they contribute to better material and product quality. The process of configuring inspection conditions can also be automated, reducing the workload placed on operators when performing inspections of high-mix, small-lot products where configuration changes are frequent. Of course, they also make sensory inspections, which in the past have been highly dependent on the skills of individual inspectors and whose results have been highly subjective, into objective inspections, while also retaining evidence for use in confirming inspection judgments and analyzing product conditions. They are continuing to evolve to meet demands for more uniform product quality and the reduction of product losses by rapidly identifying the factors that cause defects.

Let's look at a few of Toshiba's appearance inspection devices.

First is the M9100 series of web appearance inspection devices, used to inspect the surfaces of sheet products (web products) such as film, iron, and non-ferrous metals. It offers Toshiba's fast, accurate image processing technologies, and can work with diverse materials and immediately detect various defects, such as staining, holes, wrinkles, and contamination. It combines a specialized camera and lights, and is assembled to work with existing production lines. The camera scans the surfaces of sheet products and the images it produces undergo image processing to identify defects. AI is now being used to automatically categorize the defects the system finds (Fig. 1).

For customers seeking greater appearance inspection accuracy and more uniform product quality, and for customers who are struggling with passing on the skills of veteran inspectors to junior personnel, we offer AI Visual Inspection Package, which uses AI technologies. This solution uses Toshiba's own unique AI technologies to rapidly automate advanced inspection work.

The AI Image Automatic Inspection Package is made up of learning software and judgment engine software. The learning software creates a learned model ("non-defect model") using a non-defect learning method. It then applies that non-defect model, together with the judgment engine software, to images of actual items to determine if they are defect-free or defective.

The non-defect learning method is an AI learning method developed in-house by Toshiba.

AI learning performed using technologies such as deep learning requires the preparation of a large amount of learning image data for both defective and non-defective items. However, this creates a major challenge: how to collect sufficient image data to train the AI -- not just image data for non-defective items, but also image data for defective items, which occur far less frequently.

That's why Toshiba has developed a technology for training AI using image data for non-defective items alone. It eliminates the need for image data for defective items, which reduces the time and effort required to prepare image data to be used in learning. This makes it possible to created learned models in little time (Fig. 2).

Also, generally speaking, creating learned models that don't overlook defects requires narrowly defined thresholds for distinguishing them from non-defective items. However, when threshold values are too narrowly defined, non-defective items are often misidentified as defective (false detections). The non-defect learning method uses image data from non-defective items, repeatedly training the AI to learn the parameters of non-defective items. This creates a non-defect model with optimized threshold values.

A verification project was carried out at one component plant, looking at roughly 90,000 post-weld inspections. It found that 31 defects were overlooked during visual inspections by veteran inspection personnel, while the AI using the non-defect model missed only two defects.

As this example shows, this AI is capable of image diagnosis with an extremely high level of accuracy. The AI Image Automatic Inspection Package, which uses this non-defect learning method, can not only be embedded in Toshiba's own appearance inspection devices, but it can also be linked to existing customer appearance inspection devices or used via the cloud. This makes it a versatile solution for various customer inspection environments.

Toshiba has been researching and developing advanced technologies for detecting minute defects, measuring just microns, that are invisible to the naked eye, and deploying those technologies in its commercial products.

The appearance inspection devices and AI technologies we've discussed have worked by inspecting images taken with cameras to determine the quality of the photographed items. Detecting minute defects, too small to be seen with the naked eye, is therefore an urgent task for them. Believing that the evolution of optical inspection technologies was a major key to achieving more advanced quality inspections, we developed OneShotBRDF optical inspection technology.

The properties of light play an important part in optical inspection. When light strikes the surface of an object, it is reflected. In regular (or mirror) reflection, light bounces off an object at the same angle it struck the object. With diffused (or scattered) reflection, light bounces off in all kinds of directions. Diffused reflection occurs, for example, when the surface of an object is rough, or if it has pits or projections such as scratches. Here, we'll refer to light reflected via regular reflection as "regular reflection light" and light reflected via scattered reflection as "scattered light."

Normally, when performing optical inspection, the camera is positioned where it can best capture the most light -- the regular reflection light. This makes the object being inspected very easy to see, but it makes it difficult to detect minute scratches or flaws that are hard to see with the naked eye. Adjusting the lighting angle and the position of the camera to better capture scattered light reflected from scratches is no easy matter, as scattered light is reflected in all kinds of directions.

That's why Toshiba developed a technology that physically separates regular reflection light and scattered light by their color, using a multi-wavelength coaxial aperture filter placed between the camera lens and the image sensor. By simply taking a single shot of the surface of the item being inspected, it can detect minute scratches, pits, projections just microns in size, invisible to the naked eye, and make them easy to see by using color. One of the key features of this technology is that it's all done optically, so there's no need for image processing (Fig. 3).

Toshiba aims to combine the state-of-the-art AI and optical inspection technologies we've introduced with the appearance inspection device expertise it has developed through the years, helping solve the problems the manufacturing industry encounters when it conducts quality inspections, an essential process for improving product quality.

Competition in the manufacturing industry is grower ever fiercer, and product quality demands are constantly rising. Companies need to be able to identify and remove products with even minute defects too small to see with the naked eye, maintaining a stable level of quality and shipping only high quality products. The manufacturing industry is facing tremendous challenges, such as personnel shortages and how to pass on the skills of veteran personnel.

Toshiba is working to solve these problems by using its many years of experience and know-how to develop continually evolving quality inspection automation solutions.

• TOSHIBA Cyber Physical Systems