Applying Knowledge Distillation on Pre-Trained Model for Early Grapevine Detection

Abstract

The development of Artificial Intelligence has raised interesting opportunities for improved automation in smart agriculture. Smart viticulture is one of the domains that can benefit from Computer-vision tasks through field sustainability. Computer-vision solutions present additional constraints as the amount of data for good training convergence has to be complex enough to cover sufficient features from desired inputs. In this paper, we present a study to implement a grapevine detection improvement for early grapes detection and grape yield prediction whose interest in Champagne and wine companies is undeniable. Earlier yield predictions allow a better market assessment, the harvest work’s organization and help decision-making about plant management. Our goal is to carry estimations 5 to 6 weeks before the harvest. Furthermore, the grapevines growing condition and the large amount of data to process for yield estimation require an embedded device to acquire and compute deep learning inference. Thus, the grapes detection model has to be lightweight enough to run on an embedded device. These models were subsequently pre-trained on two different types of datasets and several layer depth of deep learning models to propose a pseudo-labelling Teacher-Student related Knowledge Distillation. Overall solutions proposed an improvement of 7.56%, 6.98, 8.279%, 7.934% and 13.63% for f1 score, precision, recall, mean average precision at 50 and mean average precision 50-95 respectively on BBCH77 phenological stage.