Abstract—Effectively detecting and removing inedible harvests before or after harvesting is important for many agri-workers. Recent studies have suggested diverse measures, including various robot arm-based machines for harvesting vegetables and pulling up weeds, using camera systems to detect relevant coordinates. Although some of these systems have included monitoring and identification tools for edible and inedible targets, their accuracy has not been sufficient for use. Thus, further improvements have incorporated computing into the process based on human feelings and commonsense-based thinking, which considers up-to-date technologies and determines how solutions reflect the experience of traditional agri-workers. Our focus is on Japanese small- to middle-sized farms. Thus, we developed a fine-tuning (transfer-learning)-based deep learning system that gathers field pictures and performs static visual data analyses using artificial intelligence (AI)-based computing. In this study, pictures included kiwi fruits, eggplants, and mini tomatoes in outdoor farmlands. We focused on several program-based applications with deep learning-based systems using several hidden layers. To align with this year’s technical trends, the data is presented concerning two patterns with different target layers: (1) all bonding layers with a revised pattern, and (2) some convolution layers with a visual geometry group (VGG) 16 and picture classifier created by convolutional neural network (CNN) revised pattern. Our results confirmed the utility of the fine-tuning methodologies, thus supporting other similar analyses in different academic research fields. In future, these results could assist the development of automatic agricultural harvesting systems and other high-tech agri-systems.
 

Index Terms—picture classification, deep learning, fine-tuning, Keras, Theano

Cite: Shinji Kawakura and Ryosuke Shibasaki, "Distinction of Edible and Inedible Harvests Using a Fine-Tuning-Based Deep Learning System," Journal of Advanced Agricultural Technologies, Vol. 6, No. 4, pp. 236-240, September 2019. Doi: 10.18178/joaat.6.4.236-240