Publication: Data augmented yolo algorithm-based machine vision inspection technique for wood defect detection
Loading...
Date
2023-08
Authors
Grace Tan Yih Chin
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Traditional method of detecting wood defects which is by using manpower causes problems to the industry such as low efficiency, low accuracy in defect detection, high cost, and longer time spent. In a result, deep learning which is the application of Artificial Intelligence is needed in this project so that a system is enabled to have automated studies from data and improve its accuracy from experience by using the data and algorithms of the labelled wood defect images [1]. Besides, this project is significant to improve the wood’s quality and commercial value. For this project, supervised learning is applied on the step of training datasets. For the beginning of the project, the images of the wood are supplied by the factory. After that, the defect of images needs to be labelled. A software named “labellmg” is used to label the defect of the wood. After finishing labelling, the data of the wood images are collected, and the images are augmented by using “roboflow”. After the process of image data augmentation, the environment is set by using Anaconda Navigator. Then, the datasets are trained so that the data of the labelled images can be used to train the YOLO (You Only Look Once) algorithm by using PyCharm. There are two versions of YOLO algorithms that are used to compare the results of detecting the datasets, which are YOLOv3 algorithm and YOLOv5 algorithm that consists of YOLOv5s model and
YOLOv5m model. Image data augmentation is used in all YOLO algorithms and refined transfer learning is applied in YOLOv5s and YOLOv5m models to improve the performance of datasets training. For the results, YOLOv5m model has better results than YOLOv3 and YOLOv5s models, with the average accuracy rate of 99%. The implication of the research is that deep learning algorithms such as YOLO algorithm are suitable for detecting the types of defects in wood images so that the wood’s quality
and commercial value can be enhanced