基于关系预测任务的金相组织图像识别实验系统设计与实现  

Design and implementation of experimental system for metallographic organization image recognition based on relational prediction task

作　　者：张利欣[1] 孙涵 曾慧[1] 刘红敏 ZHANG Lixin;SUN Han;ZENG Hui;LIU Hongmin(School of Intelligent Science and Technology,University of Science and Technology Beijing,Beijing 100083,China;China Telecom Digital City Technology Co.,Ltd.,Baoding 071700,China)

机构地区：[1]北京科技大学智能科学与技术学院,北京100083 [2]中电信数字城市科技有限公司,河北保定071700

出　　处：《实验技术与管理》2023年第12期16-23,共8页Experimental Technology and Management

基　　金：国家自然科学基金(62273315);教育部产学合作协同育人项目(201902020003);北京科技大学2020年本科教育教学改革项目(JG2020M24);河南省高等教育教学改革研究与实践项目(2021SJGLX260Y)。

摘　　要：为提高金相组织图像分析的自动化水平,准确获取反映材料性能的晶粒度等关键信息,提出一种基于多任务学习的晶界分割方法。该方法设计了一种关系预测辅助任务,通过关注晶粒内部信息,从而提升晶界提取的精度;同时采用对比学习训练特征提取器,使模型对晶界拓扑结构具有更强的针对性。在Pyside2框架下实现了基于关系预测任务的金相组织图像识别系统的构建。通过实验验证,本系统不仅可帮助实验室技术人员完成晶粒度等微观组织信息智能检测,且在软件的适用性、检测效率和准确性上都有大幅提升。In order to enhance the automation level of metallographic image analysis and accurately obtain key information reflecting material properties such as grain size,a grain boundary segmentation method based on multi-task learning is proposed.This method designs a relationship prediction auxiliary task to improve the accuracy of grain boundary extraction by focusing on the internal information of grains.Simultaneously,contrastive learning is used to train feature extractors,making the model more targeted towards grain boundary topological structures.A metallographic microstructure recognition system based on relationship prediction task is implemented under the Pyside2 framework.Through experimental verification,this system not only assists laboratory technicians in intelligently detecting microscopic structural information such as grain size but also significantly improves the software's applicability,detection efficiency and accuracy.

关 键 词：图像处理 金相分析 晶粒度 深度学习 系统设计 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]