OKWOR, NMASICHI GRACE (2026) FRUIT QUALITY CLASSIFICATION SYSTEM USING IMAGE PROCESSING. Other thesis, Godfrey Okoye University, Enugu.
|
Text
Okwor Nmasichi's Project .pdf Download (1MB) | Preview |
Abstract
Post-harvest fruit losses have continued to be a big agricultural problem for developing nations because manual quality inspection has been found to be unreliable due to its subjective nature, vulnerability to human fatigue, lack of scalability with increased production levels, and failure to evaluate quality condition and ripeness stage together. This paper presents the development of a Fruit Quality Classification System through image processing that could automatically identify and classify fruits in terms of their quality condition and ripeness stage using digital images. A two-model pipeline consisting of a YOLOv8n object detection algorithm and Dual-Head EfficientNet-B0 classifier was created, wherein the former detects and classifies fruit images, while the latter utilizes a multitask learning framework in determining quality (Fresh/Rotten) and ripeness stage (Unripe, Ripe, Overripe) of fruit images. Both models were fine-tuned using transfer learning approach on 13,599 annotated images acquired from Kaggle Fruits Fresh and Rotten for Classification database and divided into 10,901 training and 2,698 test sets using augmentations by Albumentation library and AdamW optimizer with cosine annealing schedule. The evaluation of the dual-head efficient net showed ideal values of the accuracy, precision, recall, and F1-score of 1.0 when assessing quality and ripeness classes. For the YOLOv8n object detector, the precision and recall scores of 1.0, as well as the mean average precision at IOU=50 and at IOU 50-95, scored 0.995 on each of the six fruit classes tested. The entire architecture was implemented as a full-stack web application consisting of a React.js client-side frontend, an Express.js server-side backend, as well as the Flask inference service for the ML algorithms, with a Supabase relational database providing JWT-authentication for the users, persistence in classification history and PDF report generation. Twelve test scenarios for the user interface were completed successfully. The findings show that the implementation of object detection and dual-attribute classification simultaneously in a web application resolves two recurring gaps from the review of literature namely joint quality and ripeness assessment, as well as deployment of a web infrastructure, thereby providing a practical and reproducible contribution to the automation of post-harvest fruit quality management in agricultural and food processing contexts.
| Item Type: | Thesis (Other) |
|---|---|
| Subjects: | Q Science > Q Science (General) |
| Divisions: | Faculty of Engineering, Science and Mathematics > School of Electronics and Computer Science |
| Depositing User: | MICHAEL MADUBUKO |
| Date Deposited: | 22 Jul 2026 10:07 |
| Last Modified: | 22 Jul 2026 10:07 |
| URI: | http://eprints.gouni.edu.ng/id/eprint/5975 |
Actions (login required)
 |
View Item |
