An edge-compute machine-vision system for measuring billet length on a mill conveyor

While the significance of billet size as an indicator of both cane and harvest quality is known, it is a metric that is time-consuming and difficult to accurately measure and report. In this study, edge-computing devices were used as the basis for a machine-vision system to obtain billet measurements from mill cane conveyors. Approximately 1000 images were obtained from two mills and used to train machine-vision models to identify unobstructed billets on the surface of the conveyor load. By aligning these detections with a depth map of the conveyor load surface, an estimate of individual billet lengths can be obtained and used to calculate an average billet length for the load. The accuracy of the system can then be verified by comparing those estimates, both individually and as load averages, to known measurements. When using manually annotated images containing marked billets of known length for parameter estimation, the improved length calculation procedure was found to provide quite accurate and precise measurements of billets, showing significantly better correlation (from r=0.712 to r=0.907) than the previously employed method. When measurements are viewed by per-load averages, the tested vision model appears to be able to track to within 2 cm of hand measurements, with camera measurements appearing to be biased towards a 17cm billet length. The system can interact continuously with external mill systems for prolonged periods as intended, and the measurement method used provides a solid method for acquiring individual billet measurements. When testing by per-load averages, the vision model appears to have had insufficient diversity in the training data, resulting in bias towards the lengths most commonly seen in that training data. Diversifying the training data with images acquired during testing is expected to shore up most of the bias seen here. If this result is achieved, the system should be able to enable further optimizations to harvest and transport by providing reliable information about billet length upon delivery at the mill.