We have seen in the news that facial recognition has a problem differentiating between some muffins and Chihuahuas. A contending muffin could be characterised as having a sandy colour bun and three dark blueberries. A contending Chihuahua would be a face shot with two dark eyes a dark nose and small sandy colour head. If you look at https://www.bbc.co.uk/news/business-48842750 and slightly defocus your eyes, you too will have a problem differentiating.
Whilst the subject matter might be amusing it could be argued that it’s a skewed result. I know of few situations where a muffin is proffered at the same head height as a Chihuahua walks on a pavement, nor do I expect a Chihuahua would be spotted on a conveyor belt at a muffin factory. Context is everything. We can increase the accuracy of AI in recognition processes if we limit the software to what it is expected to see. For instance, in a beer factory producing ‘own brand’ bottled beer for a number of different supermarkets, we can use AI recognition to differentiate between images of the beer label to direct the various bottles to the correct shipping container and then on to the corresponding supermarket. The AI here would be used to best effect to still recognise labels if they were presented to the camera at rogue angles. Spending programming time on recognising a muffin would be ridiculous.
AI is used to recognise number plates (ANPR). Number plates differ markedly between countries both in style and syntax. The time it takes to process all possible variations in plate would make recognition almost impossible on a motorway where high volumes of traffic are passing at high speed. Therefore a pragmatic approach is taken to restrict recognition algorithms to that of the home country and maybe one or two others. However, at a port where traffic moves slowly, more recognition algorithms can be employed to take account of the greater incidence of foreign plates.The AI software developer should be aware that other technologies may help considerably in improving recognition and reducing the complexity of such diverse algorithms required to differentiate between the head shot of a Chihuahua and a muffin. Use could be made of alternative lighting (maybe infrared) where a blueberry would not reflect as much light as the retina of an eye!