Generalizable semi-supervised learning method to estimate mass from sparsely annotated images
Mass flow estimation is of great importance to several industries, and it can be quite challenging to obtain accurate estimates due to limitation in expense or general infeasibility. In the context of agricultural applications, yield monitoring is a key component to precision agriculture and mass flow is the critical factor to measure. Measuring mass flow allows for field productivity analysis, cost minimization, and adjustments to machine efficiency. Methods such as volume or force-impact have been used to measure mass flow; however, these methods are limited in application and accuracy. In this work, we use deep learning to develop and test a vision system that can accurately estimate the mass of sugarcane while running in real-time on a sugarcane harvester during operation. The deep learning algorithm that is used to estimate mass flow is trained using very sparsely annotated images (semi-supervised) using only final load weights (aggregated weights over a certain period of time). The deep neural network (DNN) succeeds in capturing the mass of sugarcane accurately and surpasses older volumetric-based methods, despite highly varying lighting and material colors in the images. The deep neural network is initially trained to predict mass on laboratory data (bamboo) and then transfer learning is utilized to apply the same methods to estimate mass of sugarcane. Using a vision system with a relatively lightweight deep neural network we are able to estimate mass of bamboo with an average error of 4.5% and 5.9% for a select season of sugarcane.