Detection Of Surface Defects Of Highlight Surface Based On Deep Learning

Metal surface defect detection is an important link to ensure product quality in the manufacturing process.In the process of automobile hub production,rapid and accurate identification of hub surface defects plays an important role in improving production efficiency and ensuring product quality.As a metal aluminum material,automobile hub has the characteristics of high reflection and complex background,which has a great influence on the automatic detection of surface defects.In this paper,the deep learning object detection algorithm is used to detect the defects on the surface of the wheel hub.In the data processing stage,the generative adversarial network is used to remove the highlights on the surface of the wheel hub to enhance the defect characteristics.In the defect detection stage,the refinedet network structure is optimized,and the experiments are carried out on the non-highlight data set and the mixed data set of the wheel hub to improve the detection effect.The main research work of this paper is as follows:1.Establish the data set of hub surface defects.According to the appearance inspection standard of a domestic hub manufacturer,the common defects on the hub surface are divided into three categories:defect point,scratch and plaque.Through data cleaning and manual annotation,the defect data that accords with the reality is screened out,and the algorithm experiment of surface defect detection is carried out based on this.2.Research on the method of removing the highlights on the hub surface.The advantages and disadvantages of single view point polynomial adjustment,multi view point feature matching and generative adversarial network are compared,and the multi-scale network structure is used to improve the image quality.Through qualitative and quantitative analysis of the effect of highlight removal,we can provide reliable data samples for the detection of highlight surface based on depth learning.3.Wheel hub surface defect detection experiment based on non highlight data set.Through a large number of open source experiments on non highlight data sets,the advantages and disadvantages of the two-stage detection network represented by Faster-RCNN,the one-stage detection network represented by SSD and the refindet detection network are analyzed.By comparing the detection results of non-highlight data generated by the generative adversarial network with the actual data,it is proved that the data after the high light removal is feasible.4.The experiment of hub surface defect detection based on mixed data set.In order to simulate the actual production line image acquisition situation as much as possible,a mixed data set composed of non-highlights data set and highlights data set is used in the research.Through the sub module topology structure,the residual block width is increased,the refinedet basic network is optimized,and the NMS method is improved for the occlusion problem of overlapping detection frame,which improves the defect detection effect without increasing the complexity of network parameters.