An Effective Pistachio Classification by Ensembling Fine-tuned ResNet20 and DenseNet Models

The process of manually classifying pistachios presents various obstacles that may affect the precision, productivity, and overall efficacy of the sorting procedure. The subjective interpretation of pistachio characteristics by human assessors may result in inconsistent classification decisions. The outcome of sorting may differ due to preferences, experience, or tiredness. Manual classification of pistachios is time-consuming and hence labor-intensive activity that often calls for large numbers of human personnel. Accordingly, this proposed work intends to develop an automatic as well as a deep learning-based model by leveraging on the strengths of two architectures, ResNet21 and Densenet model; hence this study presents a new procedure for the categorization of Pistachio nuts. The investigation focuses on the impact caused by different learning speeds and their bearing in classifying pistachio fruit’ models. A high-resolution dataset comprising images belonging to two different types has been collected from Kaggle for training as well as validation purposes. A very careful procedure includes data augmentation, model adjustment, and transfer learning to enhance the model for pistachio categorization. The outcome of the experiments suggests that the hybridized version of the ResNet21/DenseNet model performs the best when trained with specific learning rates. This work contributes to precision agriculture by providing an achievable and operational technique for robot sorting pistachio fruits. When implementing the model with the learning rates of 0.01, 0.03, 0.05, and 0.07 the highest accuracy of 99.3% is reached by putting it to 0.01. In addition, the lowest has been the loss of 0.0137 when the learning rate was set to 0.01. These have yielded a precision of 0.9952, recall of 0.9904, and F1 score of 0.9921.