深度学习与合成数据集的锻模裂纹识别方法  

Crack recognition method of forging die based on deep learn⁃ing and synthetic dataset

作　　者：韩尚汝 程联军 HAN Shangru;CHENG Lianjun(College of Mechanical and Electronic Engineering,Qingdao University,Qingdao 266071,Chi-na)

机构地区：[1]青岛大学机电工程学院,山东青岛266071

出　　处：《模具工业》2025年第3期1-6,共6页Die & Mould Industry

摘　　要：使用深度学习视觉检测网络的算法进行锻模裂纹图像识别时,存在锻模裂纹缺陷训练样本不足和目标检测精度低的问题,为解决这些问题,首先使用虚拟模型和物理渲染的方法合成锻模的裂纹样本进行训练,其次在双阶段图像识别模型Faster-RCNN的基础上进行改进,将主干特征提取网络换为ResNet50加FPN的结构,ROI pooling层换为ROI Align层,在模型中加入通道注意力机制,改进后的模型在合成锻模裂纹验证集上AP、F1评分、召回率、查准率分别提高了5.27%、0.15%、1.94%、24.07%;最后使用合成数据集训练的改进模型对锻模裂纹样本进行识别,可以完成对实际裂纹图像的识别。When the deep learning visual detection network algorithm was used to recognize the crack image of the forging die,there were problems such as insufficient training samples of the crack defect and low detection accuracy of the crack target of the forging die.To solve these prob⁃lems,firstly,the method of virtual model and physical rendering were used to synthesize the crack samples of the forging die for training.Secondly,based on the two-stage image recognition model Faster-RCNN,the backbone feature extraction network was changed into the structure of ResNet50 plus FPN.The ROI pooling layer was replaced by the ROI Align layer,and the channel at⁃tention mechanism was added to the model.The AP,F1,recall and precision of the improved mod⁃el on the synthetic forging die crack validation set were increased by 5.27%,0.15%,1.94%and 24.07%,respectively.Finally,the improved model trained by the synthetic data set was used to rec⁃ognize the defect samples of the forging die,and the actual crack image could be recognized.

关 键 词：锻模裂纹识别 深度学习 物理渲染 合成图像 

分 类 号：TG315.2[金属学及工艺—金属压力加工] TP181[自动化与计算机技术—控制理论与控制工程]