改进U_Net网络的钢结构表面锈蚀图像分割方法
Improved steel structure surface rust image segmentation method for U_Net network
陈法法 1董海飞 2何向阳 3陈保家2
作者信息
- 1. 三峡大学水电机械设计与维护湖北省重点实验室 宜昌 443002;国家大坝安全工程技术研究中心 武汉 430010
- 2. 三峡大学水电机械设计与维护湖北省重点实验室 宜昌 443002
- 3. 国家大坝安全工程技术研究中心 武汉 430010
- 折叠
摘要
为实现锈蚀图像分割网络模型轻量化,同时消除非单一特征背景和锈液等类似特征背景干扰,本文将U_Net网络模型的编码部分替换为MobilenetV3_Large网络,导入基于ImageNet数据集的MobilenetV3_Large网络预训练权重,将U_Net网络模型解码部分的普通卷积替换为深度可分离残差卷积,并在上采样的过程中添加注意力导向AG模块和Dropout机制.经实验验证表明,本文设计的改进U_Net网络模型在非单一特征背景和锈液等类似特征背景干扰下,具有明显的锈蚀图像分割优势,相比于原U_Net网络模型,模型大小减少了81.18%,浮点计算量减少了98.34%,检测效率提升了3.27 倍,即从原来不足6 fps,提升至 19 fps.网络模型实现轻量化的同时,网络模型的准确率达 95.54%,相比于原U_Net网络模型提升了 5.04%.
Abstract
In order to lighten the rust image segmentation network model and eliminate the interference of non-single feature background and similar feature backgrounds such as rust liquid,this paper replaces the encoded part of the U-Net network model with the MobilenetV3_large network,imports the pre-trained weights of the MobilenetV3_large network based on the ImageNet dataset,and replaces the ordinary convolution of the decoded part of the U-Net network model with a deep separable residual convolution.And add the attention-oriented AG module and the Dropout mechanism in the process of upsampling.Experimental results demonstrate that the improved U-Net network model designed in this paper exhibits significant advantages in rust image segmentation under non-uniform feature background and similar feature background interference such as rust liquids.The model size is reduced by 81.18%compared to the original U-Net network model,resulting in a decrease of floating point calculations by 98.34%.Additionally,the detection efficiency has improved by 3.27 times,increasing from less than 6 frames/s to 19 frames/s.While the network model is lightweight,the accuracy of the network model is 95.54%,which is 5.04%higher than the original U_Net network model.
关键词
锈蚀区域分割/MobilenetV3/U_Net/注意力导向/深度可分离残差卷积Key words
rust area segmentation/MobileNetV3/U_Net/attention guided/depth separable residual convolution引用本文复制引用
基金项目
国家自然科学基金(51975324)
国家大坝安全工程技术研究中心开放基金(CX2022B06)
湖北省教育厅科研项目(B2021036)
出版年
2024
NETL
NSTL
万方数据