Anchor tuning in Faster R-CNN for measuring corn silage physical characteristics

Efficient measurement of harvested corn silage from forage harvesters can be a critical tool for farmer. Suboptimal fragmentation kernels affect milk yield dairy cows when the is used as fodder and oversized stover particles promote mould yielding bacteria during storage due to resulting air pockets. As harvester harvest hundreds tonnes per hour, an efficient robust system measuring quality in field required, however, current methods require manual errorsome separation steps or samples sent off-site laboratory. Therefore, we propose adopt Faster R-CNN with Inceptionv2 backbone detect kernel fragments images taken directly after harvesting without need separating particles. We explore strategies data sampling specialist models, transfer learning differing domains tuning anchors Region Proposal Network accommodate changes object shapes sizes. Our approach leads significant improvements average precision overlengths up 45.2% compared naive model development approach, despite challenging cluttered scenes. Additionally, our models are able predict network predictions Corn Silage Processing Score (CSPS) measure introduce chopped named Overlength (OVPS). For both scores obtain strong correlation against physically measured r2 0.66 CSPS, 0.79 0.95 OVPS at two verbal theoretical lengths cut.