Applying of Semi-supervised Learning and Data Augmentation for Detecting Citrus Fruits

Abstract

In Kochi Prefecture, Yuzu is widely cultivated, and its cultivation managements using information technology have been getting attention in the agricultural field. In order to predict shipping by automatically counting crops, location detection of fruits using CNN has been proposed. Here, there is a problem that the cost of creating supervised data for supervised learning is high. This paper proposes improvement methods using Noisy Student based semi-supervised learning with a teacher model and a student model as a way to utilize easily collected unsupervised data. We examine three improvement methods: scale expansion, pseudo-label generation using test time augmentation, and transfer learning to the student model using the weights of the teacher model. As prediction models, we use two trained models of YOLO v5. The transfer learning using only the generated teacher data results in mAP of 38.1%, while the transfer learning using scale expansion results in 70.6%.