Deep Learning for Precision Grade-Quality Classification of Pomegranates

The conventional methods of manual grading of fruits after agricultural harvest have always been laborious, time-consuming, and prone to human errors. In order to speed up the grading process without human intervention, an automatic system is the need of the hour. This research has led to the development a Convolutional Neural Network (CNN) based deep learning technique to provide real-time grading of pomegranate based on its visual features. This automatic grading technique would help farmers in minimizing their efforts and time spent on manual sorting, at the same time ensuring error-free and uniform sorting and packaging of quality fruits that would, in turn, enhance the storage time post packaging. The methodology uses nine pre-trained models to train on a custom pomegranate dataset with twelve grade-quality classes of pomegranates. The grade-quality grading is obtained using the fine-tuning approach in transfer learning, reducing the computational resource requirements and drastically improving inference speed, which is necessary for real-time model deployment scenarios. The results reveal that the fine-tuning of DenseNet-169 model outperformed the other models by achieving highest values of 99.20% and 1 for (Accuracy, Precision, Recall, and Fl-score) and AUC, respectively.