Seyed Iman Saedi - Faculty of Agriculture, Shahrood University of Technology, Shahrood, Iran

The recognition of fruit maturity stages in orchards is valuable for growers as it provides insights into the required treatments that would lead to the most efficient use of resources. Such recognition could also facilitate the development of robotic fruit harvesters, which depend on an appropriate vision-based recognition system capable of identifying only the ripe fruit. To address these issues, this study investigated the development of an image-based fruit classification model that can effectively discriminate between unripe, underripe, and ripe apricots in natural orchard conditions. The developed model utilized deep convolutional neural networks and transfer learning with VGG۱۶ network and ImageNet weights as the base architecture. The model was fine-tuned through sets of well-known layers including convolution, max-pooling, batch normalization, dropout, global average pooling, and others. Data augmentation was used to minimize the risk of overfitting. The model was trained and optimized using categorical cross-entropy loss function and Nadam optimizer. Results showed that the model effectively distinguished the three classes with a loss and accuracy of ۱.۳۵ and ۰.۷, respectively, for unseen test data. The model's performance was also evaluated using metrics such as the confusion matrix and classification report. The promising results of this study suggest that the developed model can be considered as a reliable tool in developing a fruit ripeness detection sensing system to be embedded in robotic fruit harvesters or UAVs

Apricot; Ripeness; Deep Learning; RGB images; Classification