Classification of hazelnut kernels with deep learning

Abstract

In this study a deep learning-based, highly accurate network structure was applied to discriminate the “whole kernel” of hazelnut from other classes such as “damaged kernel”, “shell”, and "undersized”, which are present in small quantities after mechanized sorting and later subject to manual sorting. An industrial setup was developed to generate datasets for these four classes, and 2094 images of each class were recorded. With the obtained datasets, EfficientNetB0- EfficientNetB1- EfficientNetB2- EfficientNetB3 and InceptionV3 network structures were first trained from scratch and the highest test accuracy was calculated as 97.85 % for the InceptionV3 network structure. To further increase in the test accuracy, the transfer learning structure was used by applying the Imagenet coefficients. As a result of the training using Imagenet weights, the highest test accuracy was calculated as 99.28 % for the EfficientNetB2 and EfficientNetB3 structures. Although both networks give the same result in the classification of four hazelnut classes, the success in discrimination of the “whole kernel” of hazelnut from other classes is a key solution to decrease economic loss due to incorrect classification. From this point of view, according to confusion matrixes, it was concluded that the EfficientNetB3 structure is four times more efficient than the EfficientNetB2 structure.