Classification of first quality fancy cashew kernels using four deep convolutional neural network models
Sriram K. Vidyarthi, Samrendra Singh, Rakhee Tiwari, Hong‐Wei Xiao, Rewa Rai
Abstract
Abstract In this study, we proposed deep convolutional neural networks (DCNNs) combined with image processing to classify cashew kernels in five categories based on the adulteration of first‐class fancy whole cashew kernels with butts and pieces. Four DCNN models, including Inception‐V3, ResNet50, VGG‐16, and a custom model were implemented, and their performances were compared using model evaluators, such as sensitivity, specificity, precision, accuracy, and F1‐score. Overall, all the models showed a high performance with a minimum accuracy of 95.1% and the training and validation data curves demonstrated a good fitting of the models. Although all the models demonstrated promising potential for cashew classification, Inception‐V3 and ResNet50 neural networks delivered the most promising outcome with the highest accuracies (each 98.4%) and F1‐scores (each 96%). Custom‐built model showed the least accuracy (95.1%) and F1‐score (87.9%). The findings of current work indicate that the developed DCNN models are capable of achieving automatic, fast, and accurate cashew classification. Practical application Currently, the sorting and grading of cashew kernel are mostly manual and time‐consuming process and the deep convolutional neural networks (DCNN) or Deep Learning implemented in this study can facilitate a speedy and automated cashew classification in cashew industry worldwide.