引入卷积块注意力模块的Attention U-Net木材表面裂纹检测方法  

Wood surface crack detection based on Attention U-Net with convolutional block attention module

作　　者：项晓扬 王明涛[1] 多化琼[1] XIANG Xiaoyang;WANG Mingtao;DUO Huaqiong(College of Materials Science and Art Design,Inner Mongolia Agricultural University,Hohhot 010018,China)

机构地区：[1]内蒙古农业大学材料科学与艺术学院,呼和浩特010018

出　　处：《林业工程学报》2024年第4期140-146,共7页Journal of Forestry Engineering

基　　金：内蒙古自治区重点研发和成果转化计划项目(2022YFDZ0031)。

摘　　要：木材缺陷会影响木材的使用价值和使用期限,其中木材表面裂纹是严重影响木材外观质量和机械强度的一种木材缺陷。对木材表面裂纹的检测可以尽快发现此类缺陷木材,或为后续处理提供依据。针对现有的人工检测和自动化检测木材表面裂纹效率低、成本高、漏检率高等问题,采用引入卷积块注意力模块(convolutional block attention module,CBAM)的Attention U-Net深度学习模型对木材表面裂纹图像进行语义分割,从而达到木材表面裂纹检测的目的。引入的CBAM模块包含通道注意力机制和空间注意力机制,分别用于捕捉通道间的依赖关系和像素级的空间关系,该模块被添加到Attention U-Net网络的编码阶段,以增加感兴趣区域的权重并抑制冗余信息。最后,通过消融试验验证了Attention U-Net中加入CBAM对分割性能的提升。采用像素准确率(PA)、类别像素准确率(CPA)、召回率(Recall)、Dice系数、交并比(IoU)和平均交并比(MIoU)等语义分割评价指标评价各模型的优劣,并确定最佳模型及其参数。在自制木材表面数据集的裂纹分割中,使用AdamW优化器引入CBAM的Attention U-Net的PA、木材裂纹Recall、木材裂纹Dice系数、木材裂纹IoU、MIoU分别比使用SGD优化器的Attention U-Net原始模型提高了0.11%,4.14%,2.96%,3.58%和1.84%。结果表明,使用AdamW优化器引入CBAM的Attention U-Net能够较好地分割背景和木材表面裂纹,区分节点、表面纹理和木材裂纹,并将节点和表面纹理分割为背景。Wood defects affect the use value and service life of wood products,among which surface crack is a type of wood defect that seriously influences the appearance quality and mechanical strength of wood components.The detection of cracks on the surface of wood can find such defective wood as soon as possible or provide some wood features for subsequent treatment.For example,wood is prone to crack during the drying process,and small cracks can be reduced or eliminated by steam spraying.The detection of small cracks can provide wood conditions for steam spray treatment to ensure the quality of dried wood.At present,manual detection of cracks on the surface of wood is inefficient,costly and has a high detection error rate,so automatic detection technology has been developed fast recently.However,the existing automatic detection methods still have many problems,such as the detection accuracy is easily affected by moisture content,CT detection is costly and complicated to operate,and traditional image segmentation detection algorithms have high complexity and low efficiency.The semantic segmentation method of deep learning in wood surface crack detection is characterized by high accuracy,high efficiency,and low cost,which can effectively detect wood surface cracks and reduce the interference of defects such as knots and other factors as surface texture.In this study,the Attention U-Net deep learning model with the introduction of convolutional block attention module(CBAM)was used to semantically segment the wood surface crack image,to achieve the purpose of wood surface crack detection.The introduced CBAM module contained a channel attention mechanism and a spatial attention mechanism,which were used to capture inter-channel dependencies and pixel-level spatial relationships,respectively.This module was added to the encoding stage of the Attention U-Net network framework to increase the weight of regions of interest and suppress redundant information.Finally,the performance improvement of adding CBAM to Attention U-Net w

关 键 词：图像处理 语义分割 木材表面裂纹检测 深度学习 U-Net模型 注意力机制 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TP391.41[自动化与计算机技术—控制科学与工程]