Greg Blackman reports on the views of panellists from AIT, MVTec, Irida Labs, and Xilinx discussing AI and machine vision in a live webinar
Surface inspection and optimising processes beyond the production line, such as helping staff with tasks, were two areas where the panel saw potential for the use of neural networks in manufacturing. But deep learning algorithms for machine vision should not be considered the Holy Grail, the panel warned.
At the Austrian Institute of Technology (AIT), researchers are working on combining classical machine vision approaches and neural networks to overcome some of the problems encountered with deep learning. Petra Thanner, senior research engineer at AIT, described work into reducing the dependency on the image source – the difference in noise levels from different cameras, for instance – by pre-processing images before applying them to a neural network. She said her team had managed to train a network with pre-processed images – even with images of different optical resolutions – and applied it to an industrial problem with good results.
The group at AIT is also investigating methods like one-class learning, a technique to spot anomalies in data consisting of images of only good parts – large datasets of defects are hard to come by in industrial inspection. It is also working on improving the quality of image data using generative adversarial networks (GANs) for data augmentation, a technique that can also help with small datasets.
One specific application example Thanner mentioned was using deep learning to detect cracks and machining marks on metallic surfaces. Olaf Munkelt, managing director of MVTec Software, also noted that deep learning can add a lot to the area of surface inspection, which has traditionally been difficult to solve using rule-based approaches.
He also pointed to optical character recognition (OCR) as an area where deep learning can improve existing technology – MVTec's Halcon software now has in-built deep learning in its OCR tool. The firm has millions of samples of industrial printed characters on which to train its neural network. 'There you can really add three percentage points [to improve the recognition rate], which makes the customer happy,' Munkelt said.
This illustrates the point that Munkelt and the other panellists made, which is that deep learning has to add value first and foremost. Quenton Hall, AI system architect at chip maker Xilinx, said he's seen a shift over the last three years in the understanding of what deep learning can accomplish. But that neural networks are best applied to areas where they can increase performance or yield, or areas where it's challenging to build a classical algorithm for a task.
MVTec is now offering training courses on deep learning to give users a better understanding of where it can and can't be applied. Plus, the firm's Innovation Day on 3 February will cover some aspects of deep learning.
Munkelt said there is an onus on suppliers of this technology to explain what neural networks can achieve and how much effort is required to develop a system. Hall also made the point that, in the machine vision market, a lot of the OEMs that are developing cameras are not necessarily the same companies that will be training and deploying the model. He said: 'It's a big question for a lot of machine vision camera suppliers: how do they enable their customer to integrate their own custom deep learning model onto these hardware platforms?'
The development effort can be considerable when working with neural networks. Vassilis Tsagaris, CEO of Irida Labs, noted that it's easy to design an initial solution with 80 per cent success, but it's difficult to move to a fully scalable solution. 'You need an infrastructure not only for training and deploying, but also for taking care of the lifecycle of the product,' he said. 'You need to put the user in the loop [for real-world deployment]; you need to define the objectives, understand how you are going to work with data – more data doesn't necessarily mean a better model … understand what type of detection is important, and have a holistic approach that will deploy and feed the model throughout the product lifecycle.' He added: 'A first prototype is easy, but going into production requires more effort.'
Thanner estimated that around 70 to 80 per cent of all the work in building a system based on neural networks revolves around data – its collection, preparation, and making sure it is without bias. 'You need to know where the data is from, you need to annotate the data, you have to generate a ground truth to train the network,' she said. 'It takes lots of effort to collect and maintain a good and valid dataset.'
Thanner added that the dataset has to be balanced and cover all scenarios with no holes, with examples of good samples in all possible situations and also all possible defects. If there are two Gaussian distributions in the dataset, and in the real world only one of these datasets is represented, then the neural network won't be able to make proper decisions for the second distribution. A neural network is only able to learn on the data, not on something it hasn't seen before.
Tsagaris agreed that data handling is 'important for faster time to market and successful implementations'. He said there are different ways to make the best use of data – like augmentation methods – but it still comes down to field data as 'what's going to drive the success in machine vision'.
Irida Labs provides AI on embedded platforms, and Tsagaris believes that in the future there will be a convergence between computer vision, AI and embedded vision. Munkelt added that in the coming years the ecosystem will be understood better, and there will be more knowledge around when to perform tasks in the cloud or on the edge, and where deep learning adds value.
Write for us
Share your experience of developing a vision system using deep learning. Please get in touch: greg.blackman@europascience.com.
