Research On Metal Surface Defect Detection Models Based On Several Convolutional Neural Networks

Metal surface defect detection problems appear widely in industrial designing and manufacturing,common solutions include artificial,traditional image processing and deep learning methods.Because deep learning methods have a strong adaptability to environment,high accuracy and high level of automation,becomes a hot research topic in recent years.In this paper,the defect determination and locating for the metal material surface,i.e.Direct Bonding Copper(DBC),based on several common convolutional neural networks are studied.For the problem of judging whether there are defects on the surface of DBC that the industry is concerned about,firstly,defect determination models with VGG16 and Res Net101 are formulated,and a series of experiments based on the original image and sub-image of the DBC are accomplished.The experimental results show that although the false positive rate is low,the missing rate is relatively high.In order to reduce the missing rate,then with Faster R-CNN,based on the above two types of datasets,by optimizing the feature extraction network,anchor box parameters,using data expansion and transfer learning strategies,two new defect determination models are constructed.The experimental results show that the model based on sub-image is better and has a lower missing rate.In order to improve time efficiency,we also build a preliminary defect determination model with YOLOv4.Furthermore,for the surface defect locating problem of DBC,with Faster R-CNN,on the basis of inheriting the determination model optimization strategies,through tuning some parameters,a defect locating model with higher accuracy is established.