Skip to main content

Making machine learning work for industry

Covision Lab has ambitions to be the leading computer vision machine learning hub in Europe for industry. Greg Blackman spoke to its CEO, Franz Tschimben

Getting the best out of advances in computer vision and machine learning is a daunting task considering how much investment is being made in the area, but Covision Lab is bringing together academic research in computer vision with an industrial base to develop applied machine learning platforms.

The goal of Covision Lab, its CEO, Franz Tschimben, told Imaging and Machine Vision Europe, is to become the leading computer vision machine learning hub in Europe for industry.

With headquarters in South Tyrol, Italy, Covision Lab is a consortium formed in 2019 from seven companies. These founding firms cover various sectors, but there is a strong manufacturing presence, which is why one of Covision Lab’s subsidiaries, Covision Quality, is focusing on manufacturing surface inspection.

Covision Quality won the start-up award recently at A3’s Automate trade show in Detroit. Covision Quality was also among the leading start-ups as part of the San Francisco Alchemist Accelerator programme, which helped raise its profile in the US.

The research side of Covision Lab is headed by Professor Oswald Lanz at the Free University of Bolzano Bozen, where partner companies – Nvidia is one – can sponsor PhDs. Tschimben noted that the research side of the company firstly pushes boundaries by publishing code for developers, and also, over the long-term, will help Covision Lab attract some of the best computer vision talent.

Covision Quality’s technology is based on unsupervised machine learning, in that the software learns what a good part looks like and deduces the defects from anomalies in the image data. It is not trained with defects, just images of good parts, although often the manufacturer’s CAD model is also used as part of the software pipeline.

Covision Lab has two offshoots, both based on the same unsupervised machine learning technology: Covision Media, which uses the algorithms for rendering sportswear as hyper-realistic 3D models for e-commerce – Gore-Tex uses the software, for example – and Covision Quality for manufacturing inspection.

Covision Quality came about through Alupress’s involvement as one of the founding companies in the consortium. Alupress makes die-cast aluminium parts for the automotive industry. The focus at the moment for Covision is detecting surface defects, mainly on metals, plastics and packaging, although Tschimben said it plans to expand to inspect other surfaces in the future.

Covision Quality is targeting mid-size manufacturing companies – 2,000 to 15,000 employees – companies that don't have the specialised personnel to deploy traditional machine vision at scale. Tschimben said: ‘[Mid-sized manufacturing firms] can have 10 per cent of production lines equipped with machine vision, but they might have difficulties to scale to more because of programming legacy – non-machine learning-based visual inspection software is a lengthy, difficult task that requires specialised engineering personnel that is often hard to find and hire. Manufacturers therefore often work with third party consulting companies. That's where we come in.’

The unsupervised learning approach means that specialised personnel isn’t needed and the software doesn’t need to be programmed. This makes visual inspection much more scalable, according to Tschimben.

One of Covision Quality’s success stories has been with GKN Sinter Metals, one of the leading sinter metal companies in the world with thousands of employees globally.

Covision Quality began by implementing its software at one GKN plant in Europe, inspecting a set of metal parts in real time at production speeds of hundreds of milliseconds per part. From this assessment, GKN took the decision to deploy Covision software across multiple production lines at sites in Italy, Germany and the USA. One of the reasons was that GKN calculated that Covision’s technology would be 20 times faster at deploying new vision systems compared to traditional methods. Tschimben said that it takes roughly a couple of hours to have its software automatically program a visual inspection system, which then can be deployed after going through various accuracy and reliability tests with the customer.

One pain point for customers using traditional vision technology, Tschimben noted, is that it's difficult to program a classical machine vision system to inspect for every defect. ‘Manufacturing companies sometimes struggle with traditional visual inspection systems and software, because they are hard coded on certain specific defects,’ he said. The advantage of Covision’s software, according to Tschimben, is that it can handle varying inspection conditions and changing parts easily. ‘In the long run we help our customers be faster and more accurate,’ he said.

The software is trained on around 100 to 200 images of mostly good parts. An operator then decides whether the predictions the software makes as to possible defects are accurate and retrains where necessary. Transfer learning can be used where the customer has similar defects, like dents, burrs, missing geometry, or changing colour. There’s also a continuous learning approach where the software learns to account for changes in the production process over time.

Along with finding defects, the system has an aggregate function to give high-level statistics on how the lines are running.

Tschimben said the system does not save large amounts of data, but that Covision installs workstations on the production site – on average, one workstation would cover four production lines – to handle the data generated by real-time requirements of running at speeds of up to one part every 200ms. Covision collaborates with Nvidia on the workstations. 

At the moment Covision Lab has two PhD students: Tsung Ming Tai, in collaboration with Nvidia, whose work is on video understanding and forecasting of actions and activities; and Cynthia Ugwu, who is focusing on anomaly detection for visual inspection. ‘Most of the PhDs are working on shaping the status quo of research and solving large industrial challenges,’ Tschimben stated. ‘In the computer vision and machine learning space, research and applications are closely linked, as this year’s CVPR 2022 has shown once again.’ Tsung Ming Tai was awarded second place for one of the challenges at the Computer Vision and Pattern Recognition (CVPR) conference. ‘This will benefit our manufacturing use cases in the long run,’ Tschimben continued. ‘We have many companies reaching out to us already in order to start research collaborations, beyond the mere use of our products.

‘Our close collaboration with research departments at universities will guarantee that we stay at the forefront of [machine learning] and continue to push the state-of-the-art ourselves,’ he added. ‘Our unsupervised machine learning approach to visual inspection is a very novel approach to machine learning… at industrial scale.’

Tschimben concluded: ‘There's a high acceptance rate for machine learning at the moment, so the time is right [for Covision to grow].’

Media Partners