|
การจำแนกความสุกของทะลายปาล์มสดด้วยการเรียนรู้เชิงลึก |
|---|---|
| รหัสดีโอไอ | |
| Creator | ฐิติรัตน์ ศิริบวรรัตนกุล |
| Title | การจำแนกความสุกของทะลายปาล์มสดด้วยการเรียนรู้เชิงลึก |
| Contributor | ปัญจวรรษ แซ่เฉิน |
| Publisher | King Mongkut’s University of Technology Thonburi |
| Publication Year | 2566 |
| Journal Title | KMUTT Research and Development Journal |
| Journal Vol. | 46 |
| Journal No. | 1 |
| Page no. | 81-104 |
| Keyword | Artificial Intelligence, Computer Vision, Deep Learning, Image Classification, Grad-CAM |
| URL Website | https://journal.kmutt.ac.th/ |
| Website title | เว็บไซต์วารสารวิจัยและพัฒนา มจธ. |
| ISSN | 2697-5521 |
| Abstract | Currently, palm-oil mills in Thailand employ staff to visually assess the ripeness of fresh palm bunches at point of purchase. This practice suffers from some limitations; the staff sometimes misjudges the ripeness of fresh palm bunches. As a result, palm-oil mills have a higher than usual cost of purchasing palm bunches. The present research therefore aimed to develop a deep-learning model that can be used to analyze photographs of oil palm fresh fruit bunches and accurately classify their ripeness. The results showed that the ResNet50(C) model provided the best adjusted accuracy at 90%, where the F1 score of each category was noted to be higher than 80% ripeness. However, it is a large model and requires a longer average testing time (405 MB, 2.48 seconds (GPU), 3.27 seconds (CPU)). If a smaller model size is desired and a faster average testing time is needed, DenseNet121 (Train from Scratch) model can instead be considered. Although the model provided the adjusted accuracy at 86%, slightly less than that of the ResNet50(C) model, its F1 score for each class was as well above 80%; the model is smaller and requires a shorter average testing time (100 MB, 1.76 seconds (GPU), 2.56 seconds (CPU)). The model is also highly robust to changes in the brightness of the photographs (range -70 to +70 from normal sunlight). |
