Abstract
Date palm (Phoenix dactylifera L.) is one of the most valuable domesticated fruit trees characterized with thousands of varieties that grow in different arid regions. Because of high diversification, discrimination between date varieties at different post-harvest handling and production stages is necessary. In this study, five different CNN (Convolutional Neural Network) models, namely ResNet18, ResNet50, MobileNet, GoogleNet and DenseNet, are used as fine-tuning tools for the classification of five Moroccan date fruit varieties: ‘Mejhoul’, ‘Boufeggous’, ‘Assiane’, ‘Aziza’ and ‘Bousthammi’. The features of MobileNet, the most successful of these CNN models, were analyzed with an RNN (Recurrent Neural Network)-based LSTM (Long short-term memory) architecture. In addition, feature selection is performed for MobileNet features to achieve a more successful classification with fewer features. As a result of LSTM-based classification of both original MobileNet features and selected features, higher classification accuracy was achieved in comparison with other CNN models. Moreover, LSTM with selected features provided the most successful discrimination ability. The accuracies obtained as a result of the classification of original MobileNet features and selected features with LSTM were 99.63% and 99.70% respectively. Overall, the results indicated that the LSTM-based architecture with fewer features improves the success of existing CNN models for date fruits.