Smart Sorters Part 1:
How Smart is Your Fruit Sorting?

"...manual fruit sorting is still the preferred method in some parts of the world"

To answer the question, let’s start by defining ‘smart’. Prior to the invention of vision fruit sorters, fruit sorting was done by people. Humans have extraordinary vision and learning capacity. They can distinguish between the stem, blossoms, wind scars, decay, and various diseases, and sort them out accordingly. Even more remarkable is that humans can be trained to sort fruit with a few examples and gradually develop their fruit sorting skills over time. Sorting volume can be easily adjusted by adding or decreasing the number of people on the sorting line. Not to mention that manual fruit sorting is still the preferred method in some parts of the world.

So, how does a vision fruit sorter measure up to manual sorting? Vision sorters, with all of their silicon, definitely have more computing power. However, computing power alone does not indicate how smart fruit sorting is, as it can have all of the world's computing power and do nothing useful. Not to mention the cost! So instead we decided to investigate how to maximize the sorter’s computing power instead of expanding it.

We assess our fruit sorting in the same way we assess a human fruit sorter: can it quickly and accurately identify various features and defects? Is it capable of sorting without extra supervision? And can it improve over time? To answer these questions, we turned to machine learning. Building a fruit sorter with machine learning is a lot like building a NASCAR; everything has to be meticulously engineered and fine-tuned to make it work. Regardless, the result was a fast vision sorter that can grade fruit at the physical limits of the fruit and conveyor.

"...while the sorter with machine learning was fast, it wasn’t smart."

But we discovered that, while the fruit sorter with machine learning was fast, it wasn’t smart. Because, like NASCAR, the sorter must be operated by operators who have undergone specialized training. It also doesn’t learn over time, can’t work without constant supervision, and like race cars, when it fails, it fails spectacularly. Compared to manual fruit sorting, it’s like taking two steps forward but one step back.

We wanted Sunsort^ai to be more like an intelligent employee, who performs its duties diligently, learns over time, works at the pace and parameters set by the packhouses without needing extra supervision. To accomplish this, we used a type of machine learning known as a deep neural network. A deep neural network, like a human, not only learns to perform the tasks at hand but also improves over time as it collects more data. So, more than a half-decade ago, we set out to integrate a deep neural network into Sunsort^ai. Today we have taken it beyond a fine-tuned machine like a NASCAR and closer to human sorters who once manned the conveyors.

"A deep neural network... not only learns to perform the tasks at hand but also improves over time..."

So how smart is your fruit sorter? Does it accelerate up to 200mph but crash spectacularly on a corner, or does it perform like your top employee every day of the year?