Accurate estimation of tuber size in large potato throughput at potato storage using machine vision and machine learning techniques

In this research work, a machine vision system to sample and measure the lengths of potato 6 tubers was developed and deployed. The proposed system generates information about tuber 7 sizes in real-time and replaces the current practice adopted by the potato industry, which 8 involves costly and labour-intensive activities such as manually sampling potato tubers and 9 passing them under a dedicated laser sensor. The machine vision system, which utilized an RGB 10 camera and a mask region-based convolutional neural network (Mask R-CNN), was able to 11 detect tubers on conveyors under a wide range of light conditions with a sampling accuracy of 12 90.74% and a mean intersection over union (mIoU) of 93%. This work also describes 7 novel 13 features to characterise fully tubers detected by the Mask-R-CNN model. These features are used 14 by a Random Forest model to sample at least one and up to five tubers in clusters that ranged 15 between 54 and 231 tubers. The proposed system showed excellent performance in the detection 16 and sampling of tubers, under both static and dynamic conditions, with a worst-case relative root 17 mean square error (nRMSE) value in estimating the long and short diameters of 6.71 and 7.83%, 18 respectively, in the densely clustered scenarios. The system leverages a graphical processing unit 19 (GPU), which allows the processing of one batch of tubers per second in field, leading to an 20 increase of 21-fold in tuber sampling ability when compared to the current industrial practice, 21 while eliminating the time and labour needed to semi-manually transport and measure tubers.