Smart Sorters Part 3:
Why do we need AI?

The short answer is that AI can organize and generalize the features of defects. But isn’t that what every machine learning algorithm does? Yes, but there are nuisances.

So how does a sorter learn? Like human learning, machine learning strives to accumulate knowledge that can be applied to solve problems. Specifically, vision sorters learn how to measure size and color and identify defects through examples. These examples are typically photographs with defects that must be sorted out.

"If we were to check every fruit for every defect profile we have at ten pockets per second, we might need a supercomputer. "

Every vision sorter uses cameras to capture images of the fruit and store them in memory with all of their details. We see these details as color pixels but the sorters store them as numbers, thus brown is a set of numbers and light brown is another. A sorter can recognize many features besides colors, shapes, and sizes. And by combining these simple features, we can create feature profiles representing each defect. For example, wind scars are light brown long rectangles, while stems are brown small circles. The profiles can get quite extensive because many diseases come in all different colors, shapes, and sizes. If we were to check every fruit for every defect profile we have at ten pockets per second, we might need a supercomputer.

This was a bottleneck. We tried to compress the profiles by categorizing them. For example, if the defect is round, then it’s either oleocellosis or canker. It was faster. However, the sorter required constant adjustments because not all cases of oleocellosis are round. So the faster we make the sorter identify defects, the less accurate it becomes, and vice versa. These were the same problems that manual sorting had.

"...we discovered that some AI can improve feature profiles dynamically, making the entire AI lighter and faster."

However, AI solves this problem. During our investigation, we discovered that some AI can improve feature profiles dynamically, making the entire AI lighter and faster. How each AI optimizes the profiles is completely different. We looked at the strengths and weaknesses of each AI. There were quite a few. Ultimately we decided to work with Google AI because of its robustness.

However Google AI didn’t work out of the box for us, but it had all the tools and building blocks we needed. Using Google AI’s tools, we built an AI optimized for Sunsort^ai by combining all the feature profiles and algorithms that we had developed prior. This was a merger of two very distinct fields: our core business in citrus sorting and Google's research into advanced AI.

Sunsort^ai shifts the entire paradigm. Sorting is no longer constrained by manpower or computing power; it's only limited by the data we have collected. We've progressed from manual sorting, moved past multi-pass vision sorters, to an AI that recognizes multiple features at once, all the while running on Sunsort^ai's industrial computers. We've found what we were looking for: a smarter sorter that can take the place of human sorters at the conveyor. But this isn’t the end of our journey, this is just the beginning.

So have you checked what is under the hoods of your sorters? Is it built on one of the most advanced AI frameworks or is it just an imitation of one?