摘要
为满足人造板表面缺陷图像分割的精度要求,提出了一种改进的UNet语义分割网络模型。在传统的UNet网络结构上将编码部分改进为残差网络ResNet50并去掉连接层与平均池化层,网络通过残差块堆叠获取更多特征的底层信息;同时在跳跃连接中嵌入聚焦注意力机制的模块,抑制干扰信息,保留有效位置信息,聚焦缺陷区域并加强学习。对4种UNet网络模型的人造板表面缺陷图像分割进行仿真比较,结果表明,融合聚焦注意力机制的残差UNet网络模型在像素准确率和平均交并比等指标上有较大提升,分割精度较高。
In order to meet the requirements of the precision of image segmentation of surface defects of wood-based panels,an improved UNet semantic segmentation network model was proposed.The coding part of the traditional UNet network was modified into residual network ResNet50,and the connection layer and average pooling layer were removed.The network was stacked with residual blocks to obtain more underlying information of features.At the same time,the module of attention focusing mechanism is embedded in the jump connection to suppress interference information,retain effective location information,focus defect location and enhance learning.The simulation comparison of image segmentation of surface defects of wood-based panels based on four UNet models shows that the residual UNet model integrating the attention focusing mechanism has been greatly improved in pixel accuracy and average intersection ratio,as well as the segmentation accuracy.
作者
张平均
翁悦
王小红
李稳稳
林艺斌
ZHANG Pingjun;WENG Yue;WANG Xiaohong;LI Wenwen;LIN Yibin(School of Electronic,Electrical Engineering and Physics,Fujian University of Technology,Fuzhou 350118,China;Zhangzhou Xinhuacheng Machinery Manufacturing Co.,Limited,Zhangzhou 363999,China)
出处
《福建工程学院学报》
CAS
2022年第4期373-377,共5页
Journal of Fujian University of Technology
基金
福建省科技创新项目(2020C0052)。
关键词
聚焦注意力机制模块
UNet
残差网络
人造板表面缺陷
图像分割
attention focusing mechanism module
UNet
residual network
surface defects of wood-based panels
image segmentation