基于优化深度学习的电动桥铸件表面瑕疵识别方法

Research on Casting Surface Defects of Electric Bridge Identification Method Based on Optimal Deep Learning

针对传统电动桥铸件瑕疵检测方法普遍存在效率低、检测精度低、人工成本高等问题,文章将优化深度学习方法应用于铸件表面瑕疵检测中,实现瑕疵自主精确检测识别。依据铸造厂待检测铸件表面特征,对铸件图像进行了前期图像预处理;同时,基于优化网络模型结构,采用残差网络(Res-Net)与特征金字塔网络(FPN)构成的骨干结构,进行全图特征提取;采用区域建议网络(RPN)生成大量特征建议区域,经非极大值抑制(NMS)处理后,分别输入全连接层与全卷积完成检测任务;运用TensorFlow深度学习框架搭建模型,并采用迁移学习提高模型的泛化能力,实验结果显示,优化后的模型整体性能优于原始模型。

Traditional electric bridgecasting defect detection methods are of low efficiency,low detection accuracy,and high labor cost.Thus the optimal deep learning method is presented for the detection of the casting surface defect of electric bridges,and it can implement automatic accurate detection and identification of defects.Image preprocessing of the castings is done based on surface feature of castings.Feature extraction is done by Res-Net and FPN based on optimal model.A large number of feature recommended areas are generated by RPN.It is inputed into FC layer and FCN after NMS,then the defect detection is completed.The model is conducted by TensorFlow and the generalization ability of the model is improved by transfer learning.The experimental result shows that the whole performance of the optimized model is superior to the original model.