As someone who (also?) builds automated industrial machinery for a living (and agrees that opening the bag would be harder than picking colors), the challenge would be in improving reliability from a typical 97% to the nigh-impossible task of getting 27,000 operations correct in a row. What will you do with small pieces of skittle shell, or broken skittles that lack a shell, or 2 skittles that are bonded together, or shreds of red bag that got into the hopper? What will you do if the dye is running low, or is contaminated, and the color fades or shifts between bags? The original article had some rather difficult to automate, subjective criteria for handling these questions.
One thing to remember in the panic over automation killing off jobs is that handling the normal case is maybe 20% of the work. Handling routine errors is 80% of the work. Handling the errors that crop up once in a thousand parts - or dozens of times in this small sample - is why running lights-out is so difficult, and why we'll always need human operators and maintenance staff who can diagnose and correct unexpected problems.
That said, I'm at least 97% confident that OP, in manually counting and marking down these 27,000 items, mis-counted or mis-marked at least once. So I'm not feeling too bad about my pessimism regarding the automated solution. Even if it can't identify if a given sample is "gross" or not, it will add 1 to a counter reliably!
Also, one test of 468 bags that shows one result does not prove anything about the average number required. You'd need to run this test many times for that, and for that you'd need automation!
I work in biotech now using microscopy to understand diseases better. Same exact problems there- so much still depends on expert humans who can diagnose and deal with all the edge cases with ease. I see all these articles about amazing results using image recognition, and while some of the results are really good, they always depend on massively cleaned-up datasets.
One thing to remember in the panic over automation killing off jobs is that handling the normal case is maybe 20% of the work. Handling routine errors is 80% of the work. Handling the errors that crop up once in a thousand parts - or dozens of times in this small sample - is why running lights-out is so difficult, and why we'll always need human operators and maintenance staff who can diagnose and correct unexpected problems.
That said, I'm at least 97% confident that OP, in manually counting and marking down these 27,000 items, mis-counted or mis-marked at least once. So I'm not feeling too bad about my pessimism regarding the automated solution. Even if it can't identify if a given sample is "gross" or not, it will add 1 to a counter reliably!
Also, one test of 468 bags that shows one result does not prove anything about the average number required. You'd need to run this test many times for that, and for that you'd need automation!