文献类型：期刊文献

英文题名：Detection of natural wood defects with large color differences based on branched network

作者：Wang, Xiao[1]

第一作者：王霄

通信作者：Wang, X[1]

机构：[1]Chinese Acad Forestry, Res Inst Wood Ind, 1 Dongxiaofu, Xiangshan Rd, Beijing 100091, Peoples R China

年份：2023

外文期刊名：MULTIMEDIA TOOLS AND APPLICATIONS

收录：;EI(收录号：20231814052176);Scopus(收录号：2-s2.0-85156174282);WOS:【SCI-EXPANDED(收录号:WOS:000982739000010)】；

基金：AcknowledgementsThis work is supported by the special fund of basic research funding of Chinese central nonprofit research institutions(CAFYBB2022ZC004-2).

语种：英文

外文关键词：Computer vision; Wood knot; Deep learning; Defect identification; Double detection; Image segmentation

摘要：Computer vision is regarded as a promising technology which can achieve automatic surface defects detection in wood industry to reduce the manual works and improve reliability. In this study, an efficient algorithm used for natural wood defects detection is proposed based on computer vision. To cope with large color differences of natural defects, branched network architecture is proposed where two networks are trained separately to deal with two extremes of the wood defects: light and dark color. To avoid repeated identification caused by the double detection of the two branches, "Inter-Class suppression" is proposed to remain the real class. Area dividing process is also introduced to implement before the detection work in order to reduce the computing cost. The whole network is constructed under the framework of Region proposal network(RPN) in order to locate the defects. Results show that the proposed model can successfully detect live and dark defects which have totally different colors. More specifically, the use of branched network will enhance the feature expression of live knot which is thought to lack of specific feature. Such a strategy improves the performance on live knot identification which cannot be easily achieved by regular deep learning method. Meanwhile, area dividing can remarkably reduce the running time because it decreases the searching area for the objects detection.