Design and implementation of a convolutional neural network to classify pecan nut cultivars in a post-harvest application

Abstract

Roughly 85%-90% of the 14 000 tons of pecan nuts produced in South Africa is exported to the international market. This makes South Africa one of the four biggest exporters of pecan nuts in the world. Market survey reports indicate that the demand for pecan nuts globally is on the rise, and for that reason, South African farmers should invest in better technology to stay competitive while keeping up with the demand. The application of convolutional neural networks (CNN) has successfully applied in various domains, and recently entered also the domain of agriculture. Although not new, recent improvements and access to better tools for image processing and data analysis problem are delivering promising results. In this research, an overview is presented of current commercial sorting technology and applications where machine learning is already being researched. The application to pecan nuts is novel in the sense that there are to the author's knowledge no other studies which applied a convolutional neural network to classify pecan nut cultivars. This study laid a foundation for future research into this field by generating a dataset of over 3000 pecan nut images of three cultivars and by determined that by making use of low-cost cameras and hardware an excellent classification accuracy of 98% could be achieved. The research implemented a transfer learning process on a VGG16 and MobileNetV2 model and compared the results of both models. Other key visual parameters, such as size and colour, are also extracted and presented for future research in the field.