Tomato Fruit Image Dataset for Deep Transfer Learning-based Defect Detection

• Main Authors: De Luna, Robert G., Dadios, Elmer P., Bandala, Argel A., Vicerra, Ryan Rhay P.
• Format: text
• Published: Animo Repository 2019
• Online Access: https://animorepository.dlsu.edu.ph/faculty_research/811; https://animorepository.dlsu.edu.ph/context/faculty_research/article/1810/type/native/viewcontent
• Institution: De La Salle University

© 2019 IEEE. Tomato is considered as one of the vegetable crops with highest demand in the Philippines. Job of the farmers does not end after harvesting since the harvested tomatoes needed to be sorted according to its size. Manual sorting is the most widely recognized strategy in sorting but is very dependent on human interpretation and thus, very prone to error. This research proposed a solution that provides sorting of tomato fruit by detection of presence of defect. The study presented the generation of image dataset for a deep learning approach detection of defects based from a single tomato fruit image. Models were implemented using OpenCV libraries and Python programming. A total of 1200 tomato images classified as no defect and with defect are gathered using the improvised image capturing box. These images are used for the training, validation, and testing of the three deep learning models namely; VGG16, InceptionV3, and ResNet50. From this, 240 images are used as testing images to assess independently the performance of the trained models using accuracy and F1-score as performance metrics. Experiment results shown that VGG16 has 95.75-95.92-98.75 training-validation-testing accuracy percentage performance, 56.38-59.24-58.33 for the InceptionV3 model, and 90.58-58.46-64.58 for the ResNet50. Comparative analysis revealed that VGG16 is the best deep learning model to be used in the detection of presence of defect in the tomato fruit based from the dataset gathered.