• Jordan Gunn

Hyperspectral Imaging in Optical Sorting: A Review

One of the most impressive developments of the last few decades within food and feed production has been that of optical sorting. This is the process of inputting either raw or processed material that contains both quality product and inferior product or contaminants, and through machine vision identifying which is which and ultimately acting on that information by diverting or removing the unwanted material.

Some common uses for optical sorting include

1) Grain

2) Coffee

3) Tree nuts

The process can be incredibly rapid, and it needs to be. Any process that doesn’t work at the speed of production wouldn’t be viable and would be abandoned.

Key among metrics used to determine if the speed is adequate is what is known as “latency”. This is the time it takes to collect information on the product as it’s moving on the conveyor and then turn that information into action. Typically this is a matter of milliseconds, but the faster the better. So how can latency be reduced?

1) Computing power

2) Reduced information

3) Better algorithms

The first two methods are tricky. If you increase the computing power you’re almost certainly increasing the cost. For every fraction of a millisecond spared the cost can skyrocket. Even if you’re willing to eat those exponential costs, this approach has its limits. Reducing the information is also tricky. Ideally you’d like to maintain if not increase the amount of actionable insight you’re gaining from analysis. That requires information. The clear answer here is better algorithms.

Better algorithms not only give you a better analysis, they allow you to reach the same statistical conclusions with less data. For instance, a normal analysis of coffee cherry ripeness in optical sorting might monitor five distinct wavelengths to make a pass/fail determination. An artificial intelligence-based algorithm for spectral analysis, by contract, may be able to make that same pass/fail determination by monitoring three wavelengths. Because this requires less information, the analysis can either require reduced computing power, or even better, can provide reduced latency given the same computing power.

Another problem present within optical sorting is the lack of chemical parameters being monitored. Color isn’t always a great indicator of ripeness, for instance. There are chemical parameters that can more reliably indicate ripeness. Likewise, it’s not possible to monitor a number of parameters at all within the visible spectrum. Some of these parameters include


-Sugar content



By adjusting the sensors on a hyperspectral imaging outfit all these parameters and more can be analyzed. This allows for a better analysis, which in turn allows for a better end-product.

For these reasons it is important that optical sorting companies upgrade their automated systems to include hyperspectral imaging and AI-algorithms.

4 views0 comments