Publication: Machine vision-based nondestructive detection of undersized wood defect using YOLOV8
| datacite.subject.fos | oecd::Engineering and technology::Electrical engineering, Electronic engineering, Information engineering::Electrical and electronic engineering | |
| dc.contributor.author | Syarifah Nur Zafirah binti Syed Salehhuddin | |
| dc.date.accessioned | 2025-05-20T02:04:59Z | |
| dc.date.available | 2025-05-20T02:04:59Z | |
| dc.date.issued | 2024-07 | |
| dc.description.abstract | Wood is a natural product and one of the most exquisite and versatile materials available to us, allowing a long-term sustainable solution for endless applications. However, several types of defects in the wood that make it unable to be used in the long-term, such as pinholes, cracks, undersized, and deadknot. The wood defects will cause the wood to lose value, both financially and in terms of benefits. The main factors are reducing wood's strength, durability, and use. Defects cause the wood to appear ruined and, in certain situations, accelerate its decomposition. To prevent these factors from happening, there are several current solutions for wood defects. One of the solutions is wood defect identification using the YOLOv5 algorithm, the YOLOv7 algorithm, and a convolutional neural network model (CNN). However, there are no undersized wood defects that have been detected using any previous solutions where the industry inspects the wood manually using a machine operator, which means the operator needs to be very experienced to check carefully. In this project, the YOLOv8 algorithm will be used to investigate defect detection on the undersized wood samples with a data image size of 640, a data batch of 32, and data epochs of 100, 200, and 300. Firstly, the captured image sample undergoes the process of wood defect detection using Roboflow. After the process is done in Roboflow, the data is downloaded in the YOLOv8 version. Then, the data is trained using the YOLOv8 algorithm in Google Colab. In Google Colab, there is the installation of the YOLOv8s model, creating a train-val split, and a command line interface for simple training, validation, and inferencing models on various tasks and versions. Next, the trained YOLOv8s model is evaluated in terms of accuracy, precision, recall, and F1-score. The graph of the F1-score, precision, recall, and training and validation performance metrics will be analyzed. As a result, the highest accuracy, precision, recall, F1-score, and mAP for undersized wood defect detection are 42.90%, 88.90%, 57.90%, 60%, and 54.90%, respectively. The faster speed during training and inferencing is 0.8 ms postprocess per image and 48.9 ms postprocess per image at the shape, respectively. The undersized wood defects could be accurately detected by the YOLOv8s model of the YOLOv8 algorithm. | |
| dc.identifier.uri | https://erepo.usm.my/handle/123456789/21737 | |
| dc.language.iso | en | |
| dc.title | Machine vision-based nondestructive detection of undersized wood defect using YOLOV8 | |
| dc.type | Resource Types::text::report::technical report | |
| dspace.entity.type | Publication | |
| oairecerif.author.affiliation | Universiti Sains Malaysia |