Efficient citrus fruit image classification via a hybrid hierarchical CNN and transfer learning framework

Fruit quality assessment is paramount in the food industry and significantly influences storage conditions and duration. Classifying citrus fruits early is cru- cial for all agricultural products since it can affect market needs and result in potential financial losses. Employing non-destructive X-ray techniques to assess mandarin orange quality before reaching consumers helps maintain trust and sat- isfaction by delivering high-quality fruits. Citrus fruits rank among the world's top five consumed fruits. This study employed X-ray CT scanning of 280 cit- rus fruits (mandarine orange) stored under real-time established ambient and refrigeration conditions to perform a non-destructive evaluation of citrus fruits. The images' raw data files (in digital radiography format) were converted to JPEG using NanoVoxel software and trained on benchmark datasets, success- fully classifying images based on the storage type and period. In the proposed architecture, hierarchical fruit classification was performed by embedding CNN blocks with transfer learning models like VGG16, ResNet50, EfficientNetB0, InceptionV3, DenseNet201, MobileNetV2, and InceptionResNetV2. The VGG- Deep CNN demonstrated exceptional proficiency in classifying citrus fruit images compared to all other parallel models. Results from the suggested study indi- cated high accuracy, precision, recall, F1, and AUC scores of 0.98073, 0.98343, 0.98071, 0.97794, and 0.9994, respectively, for the training and validation set. It corroborates the testament that the classification of citrus fruits is authorita- tive, promising, and adept. Moreover, the potential investigation exhibits a 15% improvement over state-of-the-art approaches, suggesting its potential implemen- tation on a large scale in the food industry for the measurement of fruits' different external and internal characteristics.