Anchor tuning in Faster R-CNN for measuring corn silage physical characteristics
2021
Rasmussen, Christoffer Bøgelund | Kirk, Kristian | Moeslund, Thomas B.
Efficient measurement of harvested corn silage from forage harvesters can be a critical tool for a farmer. Suboptimal fragmentation of kernels can affect milk yield from dairy cows when the silage is used as fodder and oversized stover particles can promote mould yielding bacteria during storage due to resulting air pockets. As a forage harvester can harvest hundreds of tonnes per hour, an efficient and robust system for measuring quality in the field is required, however, current methods require manual errorsome separation steps or for samples to be sent to an off-site laboratory. Therefore, we propose to adopt Faster R-CNN with an Inceptionv2 backbone to detect kernel fragments and oversized particles in images of corn silage taken directly after harvesting without the need for separating particles. We explore strategies of data sampling for specialist models, transfer learning from differing domains and tuning the anchors in the Region Proposal Network to accommodate for changes in object shapes and sizes. Our approach leads to significant improvements in average precision for kernel fragmentation and stover overlengths of up to 45.2% compared to a naive model development approach, despite the challenging cluttered scenes. Additionally, our models are able to predict quality for network predictions with the Corn Silage Processing Score (CSPS) for kernel fragmentation and a measure we introduce for chopped stover named Overlength Processing Score (OVPS). For both scores we obtain a strong correlation against physically measured samples with an r² of 0.66 for CSPS, 0.79 and 0.95 for OVPS at two verbal theoretical lengths of cut.
Show more [+] Less [-]AGROVOC Keywords
Bibliographic information
This bibliographic record has been provided by National Agricultural Library