How we accidentally built the most accurate AI insect detection app

While our phone clip-on microscopes designed for kids created great interest in exploring the microscopic world, we failed to create significant revenue. So we started looking for new applications.

It was clear to us that the ever-increasing resolution of phone cameras would soon give microscopes a run and that AI would soon transform microscopy. Phones came packed with all the ingredients needed for AI implementation. Perhaps there would be a demand for AI-powered phone attachable microscopes? We were looking for a problem to solve.

Around that time, South Australia faced a regular crisis - fruit flies. Unlike much of the rest of Australia, it was a fruit fly-free state with a sizable horticultural industry and wished to remain fruit fly free. Border fruit inspections, a state-wide network of fly traps and the occasional outbreak requiring heavy and costly intervention. My co-founder Jarrad Law was keen to give our newly minted AI skills a shot at identifying the fruit flies. The fruit fly detection app was to be our killer app.

Fruit flies don't pose.

We found it difficult to position the fruit flies to take a good shot, so we 3D printed a holder with a dip to place the fly. We needed to align the insect with the lens axis. So we decided to print a ping pong ball-like shape - into which we could put the fly and gently it until it fell into the position we needed. That worked!

We did not need thousands of images.

Typically, 1000s of images are needed to build AI applications with some decent level of positive identification. We were shocked when we found that the trial application that we built worked quite nicely with just 50 images - worked quite nicely. We thought we had done something wrong; an 80% accuracy is hard to come by even with thousands of images. We repeated the experiment many times, and the accuracy did not slip. We had no idea why.

It took us a couple of months to understand why.

We could not figure it out. We kept peering at the images and finally, it occurred to us that in our spherical lighting cage, we had inadvertently removed all shadows. Also, the background was just one colour. Perhaps unlike others, we were feeding the AI application with clean data for training and providing the image for analysis.

Junk in > Junk out

Perhaps the magic was in the cleanliness of our data? So we tried a few experiments. We introduced shadows, we introduced backgrounds, and yes - the accuracy did drop. We realised that we had accidentally come up with super cool technology. With a small investment from our scientist chairman, we managed to file a patent for this game-changing technology with a very able patent attorney. Patent searchers and literature review confirmed that this accidental discovery was entirely ours.

We soon built a series of AI detection apps with just 50 training images detecting pests, plant disease and assessing prodcue quality > demo.gomicro.co

We ain't experts

It was clear to us that no one was going to take our discovery seriously. We did not a "strong track record" of publishing papers or winning industry contracts. To our surprise, the people in charge of the fruit fly prevention activity saw no use for this technology. We applied for a government matching grant - but were turned down because our microscope could not look at anything bigger than 10 mm (although we were not sure why anyone would need a microscope to look at anything bigger than 10 mm). We started sinking in a sea of disbelief. So we had only one way out.

Roll it out

So we picked a problem—a big one. The fall armyworm affects more than 300 crop types and is considered the coronavirus of agriculture. Despite the fall armyworm being detected in Feb 2020 in Australia, with warnings going out, we saw very little interest in the practical detection solutions that we had to offer. We decided to work with anyone who would work with us to collect images and validate our technology. So we made an open call to researchers through the fall armyworm collaboration portal. Thanks to enthusiastic support from Geoffrey Nyapom it generated good interest particularly from Africa - where the fall armyworm is having a devastating effect on the livelihood of over 600 million farmers. Our GoMicro Examine AI app for pest detection is now being validated in Kenya - where we had the good fortune to connected with people who wanted to do something about it,

The jury is out

We can't wait for the results. We are now working with NGOs, Departments of Agriculture and universities in Kenya, Ivory Coast, Sri Lanka, India and UK to validate our high accuracies independently. If we succeed, we will have the means to provide all farmers worldwide agronomist quality diagnostics of pests, plant disease, and mineral deficiency detection for free.