摘要
晶粒度评级精度高度依赖于准确的晶粒尺寸与形状表征,而晶界分割是界定晶粒范围的关键预处理步骤。针对铜合金显微图像中晶界对比度低、边缘模糊导致的检测困难,以及现有高精度分割算法参数量大、计算复杂度高、难以满足工业实时检测需求等问题,本文提出一种基于MobileNetV2的轻量化U-Net改进方法。通过将MobileNetV2作为主干网络解决特征丢失问题,并引入集成深度可分离卷积的ASPP模块,有效增强了多尺度语义特征提取能力。实验结果表明,改进后的模型在保持轻量化的同时,在晶界分割任务中取得了mIOU 87.66%、精确率93.50%、平均像素准确率92.79%的优异性能,显著优于传统U-Net模型,为工业现场实时晶界识别提供了可靠解决方案。
The accuracy of grain size grading is highly dependent on the precise characterization of grain size and shape,and grain boundary segmentation,as a critical preprocessing step for defining grain boundaries,directly affects the final grading result.Aiming at the detection difficulties caused by low contrast and blurred edges of grain boundaries in copper alloy micrographs,as well as the problems of existing high-precision segmentation algorithms such as large parameter quantities,high computational complexity,and difficulty in meeting the needs of industrial real-time detection,this study intends to propose a grain boundary identification method with both high precision and lightweight characteristics.This method needs to solve the problem of poor adaptability of traditional algorithms in complex micrographs,break through the real-time bottleneck in industrial scenarios,and provide reliable technical support for real-time on-site detection of copper alloy grain size.It is especially suitable for application environments with limited resources such as edge computing devices.This paper proposes a lightweight U-Net improvement method(Improved U-Net)based on MobileNetV2.The specific improvements include:Backbone Network Optimization:MobileNetV2 is used to replace the encoder of the traditional U-Net.By virtue of its inverted residual structure and linear bottleneck design,it avoids feature loss caused by feature map cropping.At the same time,it significantly reduces the number of model parameters and computational load through depthwise separable convolution thereby achieving a lightweight design.Multi-Scale Feature Enhancement:An ASPP module integrated with depthwise separable convolution is introduced into the deep layer of the encoder.It extracts local details,multi-scale context,and global information in parallel through multiple groups of atrous convolutions with different rates and global pooling,thereby enhancing the ability to capture multi-scale morphological features of grain boundaries and making up for the possible lack of feature information caused by the lightweight design.In the experiment,1080 copper alloy metallographic images(with a resolution of 512×512)were used as the dataset,which were expanded to 4320 images through data augmentation such as rotation and mirror flipping,and were divided into training set,validation set,and test set in a ratio of 7∶2∶1.Indicators such as mIOU,precision,MPA,GFLOPS,and parameter quantity were used for comparative verification with HRNetV2,DeepLabV3+,and the traditional U-Net model.The experimental results show that the improved model performs significantly better than the comparative models:Segmentation Precision:The improved U-Net achieved mIOU of 87.66%,precision of 93.50%,and MPA of 92.79%,which were 10.42%,4.40%,and 5.96%higher than those of the traditional U-Net,respectively.It also significantly outperformed HRNetV2(mIOU 74.01%)and DeepLabV3+(mIOU 74.57%).Lightweight Characteristics:The GFLOPS and parameter quantity of the model were 47.42G and 5.46M,respectively,which were 75.1%and 59.2%lower than those of the traditional U-Net(190.34G,13.39M),greatly reducing computational complexity and resource requirements.Stepwise Verification:When only replacing the backbone with MobileNetV2,the mIOU increased by 9.41%and the parameter quantity decreased by 56.5%.After adding the ASPP module with depthwise separable convolution,the mIOU further increased by 1.01%and the computational load decreased by another 25.2%,verifying the effectiveness of each improved module.The Improved U-Net proposed in this study,through the collaborative optimization of the MobileNetV2 backbone network and the ASPP module with depthwise separable convolution,not only solves the problem of feature loss in the traditional U-Net but also enhances the ability of multi-scale semantic feature extraction,achieving a dual improvement in segmentation precision and lightweight performance.With a computational cost of 47.42G FLOPs and 5.46M parameters,the model achieves an mIOU of 87.66%,and the grain boundaries in the segmentation results are more continuous with fewer breakpoints,which can meet the needs of real-time on-site detection in industry.It thus provides a reliable technical solution,suitable for edge computing devices,that enables high-precision and efficient automatic grain size grading.
作者
靖青秀
刘卫辉
常琪琪
谢伟滨
张志聪
吴瑞洋
黄晓东
JING Qingxiu;LIU Weihui;CHANG Qiqi;XIE Weibin;ZHANG Zhicong;WU Ruiyang;HUANG Xiaodong(School of Metallurgical Engineering,Jiangxi University of Science and Technology,Ganzhou 341000,China;Shanxi Huaxing Aluminum Co.,Ltd.,Xingxian 035300,China;School of Economics and Management,Jiangxi University of Science and Technology,Ganzhou 341000,China)
出处
《有色金属(中英文)》
北大核心
2026年第2期198-206,共9页
Nonferrous Metals
基金
国家自然科学基金资助项目(52461007)。
