Wood broken defect detection with laser profilometer based on Bi-LSTM network

Abstract

Detecting wood broken defects through machine vision is challenging due to the similar appearance of defect and defect-free regions on images. Laser profilometer is a reasonable solution, nevertheless, imperfect point cloud representation, such as slope profile, incontinuity of tiny defects and similarity between broken defects and sound area, poses obstacles. To overcome these challenges, this study proposes a multi-line detection method based on bidirectional long- and short-term memory network (Bi-LSTM) for real-time wood broken defect detection. The feature that represents the extent of surface damage in line-level is designed by residual extraction and sorting operation. The Bi-LSTM combines adjacent information to exaggerate semantic information of detection line. Context information extracted by Bi-LSTM are concatenated for multi-line detection to reduce computation complexity. Finally, detection results are modified by considering the information of adjacent lines of point cloud. Experimental results show that the proposed method achieves real-time detection with high accuracy.