Machine vision challenges and opportunities

In the food industry ¬– and specifically fresh produce or bakery – no two products are identical, so there is no unique and guaranteed reference point, which is the basis for all vision systems. Therefore, the biggest challenge for machine vision in the food industry is looking at the produce and recognising certain features and applying a level of intelligence that can make an assumption about something, with a very high level of repeatability. The key to achieving this is the application of artificial intelligence (AI) as vision systems can be taught to work with variability without it impacting on inspection performance or accuracy. By applying AI to machine vision it is possible to get a very high level of repeatability which is the key to high yield and minimal wastage. 

We were an early provider of 3D machine vision – a technology that is primarily understood in automation to offer depth sensing and post detection for robot picking. But it offers much more than that. We create a profile of a foodstuff in 3D and analyse it for the reference features described above. So 3D is top of the list, followed closely by the use of Convolutional Neural Networks (AI to the masses) to enhance feature extraction. 

Neural networks are the machine equivalent of brain neuron networks. Just as neurons transmit signals and information to different parts of the brain, the neural network uses interconnected nodes to teach computers how to process images. This is what is termed ‘deep learning’. If you give an AI-optimised vision system an approximation of what you are looking for by showing it examples, the neural network makes the connections. Roughly speaking, it is an elegant method of pattern matching. With 3D machine vision alone up to around 80% reliability is achievable; when you overlay that with AI, you can achieve close to 100%. 

Employing AI in fish, seafood and vegetable processing has given the ability to provide a high level of repeatability with subject matter that does not conform. Until recently, picking up a vegetable and manipulating it with two hands whilst looking at it to make a decision on what to do with it was the domain of the human. Now, those barriers are being overcome with 3D+AI and a good robot.

Machine vision technology complements rather than competes with X-ray inspection because when a vision system looks at a product it sees it differently to how an X-ray system sees it. Take the example of an apple flow wrapping line where the producer wants to check, as the bags exit the flow wrapper, that each one contains six apples. Machine vision might struggle with this application if there is a lot of print on the packaging, whereas an X-ray will see through the packaging and easily identify whether there are six apples present in the pack. 

There are a number of areas where there may be benefits. One clear benefit that has become very apparent post-pandemic is the ability to process food without it being manually handled. People often assume that machine vision is primarily used in the food industry for inspection purposes, but actually, with the current labour crisis, the big opportunity for vision lies in automation applications.

A I think a verifiable track record of the potential provider is key, as is a very detailed requirement specification and a realistic expectation of cost. I have seen many examples where installations have not met expectations – either through a vague specification document or an underestimate of the cost to develop something.