Google Visual Inspection AI helps detect defects and errors in production pipelines.
Google recently presented Vision Inspection AI, the company’s latest cutting-edge improvement in manufacturing quality control.
In June 2021, Google enriched its Google Cloud Platform with Visual Inspection AI, an artificial intelligence (AI)-driven purpose-built solution designed to quickly and accurately detect defects and errors in a variety of production pipelines. It is a continuation of Google’s previous efforts to capitalize on the manufacturing industry, which the company has recognized as one of the key target industries for its growth. Strong demand for improvements comes especially from high-tech industries, which are wasting resources and money trying to reduce faults and errors in the production cycle.
Some of these companies are already using machine learning solutions, including Google’s AutoML, to tackle this problem, but visual inspection proved to be an especially demanding task with much potential for improvement. Google published some of the results from its studies, focusing on the electronics industry, carmakers, and consumer packaged goods manufacturers, citing potential savings in production from tens of millions to even hundreds of millions of dollars yearly.
Problem Solver: Artificial Intelligence
The key innovation being introduced in manufacturing quality control is Google’s new AI system coupled with the latest advancements in image processing technology. AI systems are not completely new in the industry. Some AI applications, like Alexa, Amazon’s virtual assistant AI, are already being used by millions of customers around the world. Today, AI is a broad term used most often to describe computer-based systems that can solve complex tasks that normally require human intelligence. These tasks are still largely performed by human workers, and now, AI technology is gradually being adopted across industries.
AI systems currently have a great reputation for achieving incredible results. For example, in 2017, the DeepMind team led by Google engineers trained AlphaZero AI to achieve superhuman performance in a game of chess in only a few hours. Chess remains one of the most complex and intriguing games invented. It (still) cannot be solved by brute computational methods and requires a more intelligent approach. In a way, chess is a perfect playground for training AI systems and measuring their learning performance. AlphaZero AI proved its worth by beating the best non-AI computer engine at the time, and the team published several played games in a scientific article that shook the chess community.
The article led to significant improvements in chess engine performance over the last three years. Variants of this AI were then used for more progressively complex projects, one example being protein folding in medicine. An AI named AlphaFold is helping to improve our fundamental understanding of biology, find cures, and design drugs for the most resilient diseases. And this is just the beginning because AI applications are expected to grow exponentially in the coming years.
Traditional Methods and Modern Approaches
Visual inspection is traditionally performed by human workers who manually inspect each unit or product, one at a time, looking for imperfections. This is often slow and painful work, and workers inevitably get tired and make errors. Furthermore, visual inspection of certain products can be very demanding, for example, in the case of products with larger numbers of tiny components, or detection of subtle color differences. The human eye has not evolved to detect such details well.
For this reason, modern manufacturers often use computer-assisted machine learning (ML) solutions for visual inspection tasks. As mentioned, Google has offered its AutoML solution to manufacturers for some time. ML systems bring increased accuracy and efficiency over manual work, but they also have their limitations and disadvantages. Some need to be programmed for specific tasks, which can be a serious bottleneck. Once installed, a system needs to be loaded with a large number of images to train the system, with Google citing thousands of images. This naturally takes time, and images may not be readily available, potentially making the procedure worse. The imaging technology used and the robustness of the system may also be more limited. If the product changes or more details need to be detected—even on the same images—datasets need to be updated and the whole cumbersome procedure repeated.
The new Visual Inspection AI attempts to overshadow ML solutions, first by improving on their weaknesses and then by introducing more functionalities to ease the whole operating procedure. The most notable improvements include easier adaptability to a variety of production use cases, much faster AI training, improved accuracy and visual detection capabilities, and, crucially, no real AI expertise required to operate the AI.
How Does It Work?
Visual Inspection AI works by training the system with images: engineers in factories upload datasets of images marked proper and faulty to the system through a simple application interface. These images are then used by the AI for learning and comparison, with the exact methods hidden inside Google’s algorithm. Engineers have to mark images as “pass” or “fail” as well as label areas on the images to be checked and categorized. Multiple areas can be labeled, which enables the AI to report what kind of defect is in question. These areas can be extremely tiny. A perfect example is the electronic PCB board, which has details that are nearly impossible for humans to inspect properly. Images of up to 100-megapixel resolution are supported, guaranteeing excellent accuracy for the most demanding use cases.
The AI is then trained to recognize defects on a production line. The production pipeline equipped with cameras captures images of newly manufactured products and sends these to the AI for visual inspection. Google presents its new product effectively and created a demo that clearly displays the procedure. To watch the demo, check the link here.
Notable Advantages
According to the Google Cloud webpage, field tests performed by Google Cloud customers showed that Visual Inspection AI can use as few as 10 human-labeled images to generate accurate results. On average, it needed 300x fewer images than its ML counterparts (which often used thousands of images) and it produced up to 10 times better accuracy. Not only is it faster and more accurate, but as seen in the PCB example, the AI can detect multiple tiny components on products and classify different kinds of defects. Improved classification is an important step in making the handling of faulty units more efficient.
Adaptability and ease of use are two strong points of new technology. Visual Inspection AI can be quickly applied to vastly different industries as long as it is provided with a proper dataset of “pass” and “fail” images. The key advantages here are that the initial datasets are not large and that the AI can achieve good accuracy quickly. All of this would perhaps not be enough to attract companies to mass adopt new technology unless the system was easy to operate. Therefore, Google worked hard to simplify operation of the system to the point where no AI expertise is needed by operators. This was demonstrated practically through the company’s official demo, and there is even an option to try out Google Cloud Visual Inspection for free here.
Some other advantages include Google Cloud support, its tools, and the ability to use the solution “on-edge” without cloud capabilities. This is also a sensible addition to offer because security in factories can be of vital importance.
What the Future Holds
We can likely expect the shift to Visual Inspection AI technology over the coming years as companies battle for every edge they can get. Google has already announced that some of its large AutoML customers are eager to adopt new technology. These kinds of advancements also tend to inspire others, and competitors are likely to jump on the bandwagon by offering alternatives. Then, we can likely expect even wider and cheaper AI applications. On the topic of prices, it is interesting to note that Google is offering analysis of the first thousands of units per month for free while charging the rest per unit analyzed. Discounts are naturally larger if more units are analyzed. This business model is interesting because it can attract both small and large manufacturers, giving more potential for the success of the technology. Did Google hit a win-win scenario with Visual Inspection AI? It is certainly possible, and we are eager to see what the future holds.