基于卷积神经网络的起重机金属结构缺陷检测

Detection of metal structure defects of crane based on convolutional neural network

文中针对起重机金属结构在服役期可能产生机械磨损、疲劳裂纹等安全性问题,将视觉检测技术GDR-Net与并行卷积神经网络相结合,提出一种新的叠加卷积神经网络起重机金属结构缺陷检测法。以分次训练方式先对浅层次网络进行训练,再以GDR-Net对训练后的浅层次网络参数进行校正,最后将所提方法用于起重机金属结构缺陷检测工作中,并与现有的并行网络模型机械视觉识别方案进行了对比,结果表明提出算法结果更加准确。

Considering the safety problems such as mechanical wear,fatigue crack,etc.of the crane's metal structure in service,a new defect detection method of the crane's metal structure based on superposition convolution neural network is proposed by combining the visual detection technology GDR-Net and parallel convolution neural network.First,the shallow layer network was trained by training in stages,and then the parameters of the trained shallow layer network were corrected by GDR-Net.Finally,the method proposed in this paper was applied to the metal structure defect detection of cranes,and compared with the existing vision recognition scheme of parallel network model machines.Results show that this method is more accurate.