Designing a Hardware System to separate Defective Pistachios From Healthy Ones Using Deep Neural Networks

The aim of this study is to develop imaging algorithms to improve the grade of nuts with shell defects such as oily stains, dark stains, adhering hull, damage seed defects, and fungal decay. All these defects indicate the risk of Aflatoxin contamination. Convolutional Neural Networks (CNNs) have become prominent in various fields of machine vision and image classification. In this study, a laboratory hardware setup based on a convolutional neural network is designed for sorting pistachios. The total number of collected data is 958 images, which includes 276 images of defective pistachios and 682 images of healthy pistachios. The classification of healthy and defective images has been accomplished by 3 types of deep convolutional neural networks including Google net, resnet18 and vgg16. The accuracy and specificity of the results obtained using the pre-trained deep neural network models of Google net, resnet18 and vgg16 are 95.8% -97.1%, 97.2% -96.7%, and 95.83% -97.08%, respectively.