Abstract
An automated tropical wood species recognition system known as KenalKayu has been developed by the Centre for Artificial Intelligence & Robotics (CAIRO) to identify the tropical wood species. The system works very well in offline mode with an accuracy rate of up to 98%. But when it comes to real-time testing, the accuracy rate dropped by about 62%, partly due to low image quality. The system was trained by using ideal quality of wood images that are stored in the database. However, during real-time testing, the quality of wood image captured might be degraded due to motion blur, out of focus and illumination. Therefore, it is challenging to perform accurate recognition via real-time approach. This research proposed an improved KenalKayu prototype by using You Only Look Once version 3 (YOLOv3) algorithm to detect and classify tropical wood species via real-time approach. 60 images from 10 tropical wood species have been trained while another 60 images have been captured and tested during real-time testing. The preliminary test shows promising results where the system is now able to classify tropical wood species in real-time mode with accuracy rate for both training set and testing set are 100%. The average accuracy rate for output probability generated by YOLOv3 is 95.63%.
