Date Fruit Classification Using Transfer Learning Techniques

Abstract

Date fruit industry is considered among the important sectors in Saudi Arabia due to what this particular fruit represents to Saudi citizens in terms of cultural heritage significance. Among the several challenges that this industry is facing is the automation of the sorting process of dates which do exists in a wide variety of types, colors, size, taste, shapes, etc. This paper proposes developing an automatic classification model capable of accurately identifying different types of date fruits and providing detailed information about each type. This initial phase of this work consists of creating a large dataset of 15 types of date fruits by taking high-quality pictures, particularly from markets in Saudi Arabia. These types include the following: Safry, Mabroom almadeena, Sokary, Ajwat almadeena, Sagie, Khlas, Mjdwal, Ratab, Lobana, Rotana, Eedyah, Safawy, Khodhry, Rshodya, and Anbar. In the second phase of this resreach, three different CNN models (AlexNet, GoogleNet, VGG16) based on Transfer Learning were considered with Non-Freeze weights. Along with these models, two levels of fine-tuning were considered: (1) freezing all feature extraction layers and unfreezing the fully connected layers, where classification choices are determined; and (2) freezing beginning feature extraction layers and unfreezing the latter feature extraction and completely linked layers. Performance results affirm the efficacy of transfer learning in fruit classification tasks. The deep-structured VGG16 model, when fine-tuned with frozen weights, demonstrates exceptional validation accuracy and excellent scores in precision, recall, and F1 measure reaching 98.19%, 97.82%, 98.03%, and 98.10%, respectively.