I work in a factory that processes beets into sugar. Occasionally, conditions in the beet handling machinery cause too much water (from various possible sources) to enter the hoppers feeding the slicers, resulting in an inaccurate tare, complicating yield calculations.
While examining the end of the conveyor where the beets fall into the hoppers, it occurred to me that machine vision might be employed to detect the presence and volume of excess water by performing statistical analysis of specular highlights presented by reflections from adequately lit droplets.
This particular application would benefit from the use of neural processing, but falls outside of the “usual” cases, since the system wouldn’t be engineered for identification/classification, measurement, or counting.
My instinct tells me that accomplishing my goal should place less demand on the system than functions like classification cascades, and I could likely enjoy a relaxed frame rate and monochrome processing, further easing system loads, but I don’t know how much, if any, real-time processing would be necessary (for detecting events like “splash”), which I’m hoping won’t be necessary as this would obviously require processing at a far greater frame rate.
I’m wondering if I could construct the system to perform frame-grabs, correct for ambient light, identify specular highlights, perform statistical analysis on perceived volume averaged over a few frames, and return data representing relative volume based on droplet size (mean specular intensity), and relative count (perceived point volume), compared to a data set representing “ideal wetness”. This data would initially be used to raise alerts, but if it could be tuned to calculate actual volume with reasonable accuracy, it could be subtracted from the excess tare, while triggering other parts of the DCS to find the source.
Does anyone think this kind of “fuzzy logic” could work for my application, and if so, could I expect the system to “reckon” the volume accurately enough to use in tare adjustments, but regardless, where would I begin in selecting the appropriate platform and software?
Thanks in advance!
