Fresh fruits selection recommender

• Main Author: Loh, Zhan Herng
• Format: Final Year Project / Dissertation / Thesis
• Published: 2020
• Subjects: Q Science (General)
• Online Access: http://eprints.utar.edu.my/3903/1/16ACB02230_FYP.pdf; http://eprints.utar.edu.my/3903/

Fresh produces (fruits) quality prediction was the main idea of this project and it was hoped to serve as a fresh fruit selection recommender for common or industrial usage. With the difficulties of predicting the actual condition of the fruits through observing its external appearance as well as the internal quality factors, believing that Computer Vision could help us to solve the problem. In this project, the fruit domains involved were apple, banana and orange. Generally, the development mainly split into two different tasks or phases, fruit classification and fruit detection or localization. To identify and localize the fruit presented in one image or frame, another set of data image was collected to make annotations manually and these data were prepared to “feed” into the object detection API, which Faster R-CNN object detection model was configured as the training pipeline. A frozen inference graph was trained for fruit detection usage. Besides, to make predictions on the freshness states of these fruits, data images of two distinguishable freshness states of ‘Fresh’ and ‘Rotten’ was collected. Convolutional Neural Network (CNN) architecture was used to construct and train for classification model. In order to achieve desirable performance from the model, evaluations and analysis were made so as to make incremental improvements through using regularization method as well as transfer learning approach. To conclude the achievements of both model, the Faster R-CNN detection inference graph was welltrained as it identifies classes at most of the situations, as for the CNN classification model, it achieved the overall accuracy of 99.81% because of leveraging the pre-trained weights from VGG-16 model. Till current stage, the prototype was usable for limited fruit domains quality prediction. It was believed that it could be deployed for real-world usage if improvements and extended development were made on this system prototype.