Now,image pattern matching is used in target detection,video tracking,positioning and other fields.Template matching,as commonly used method of image pattern matching,has been focued on scholars.As the geometric feature of BIW,hole plays important role in assembly process,so the position measurement of hole is significance.But,when matching hole,they often disturbed by stains and light,so they can not obtain accurate matching results.Therefore,this paper takes the key geometric features of BIW as the application background,and combines deep learning and image restoration to match the body smooth holes and threaded holes.The main research is as follows:1)In order to solve hole training data set leads to the low model trained by deep learning,an improved network method is proposed to expand the data set.Through the analysis of GAN network and DCGAN network structure and some improvement of DCGAN,the improved DCGAN network is used to expand the image.The experimental comparison shows that the image generated by the improved DCGAN network has better quality,the extended deep learning model is 1.3% higher than before expansion.2)In order to solve the problems of stains affecting matching and poor recognition effect of strong light,raise an improved Fast RCNN.By analyzing the original network,the network is improved according to the shortcomings of the original model.Experiments show average accuracy of the improved model for the identification of stains and strong light is 91.9%,which is 3.7% higher than the original model.3)In order to solve the influence of stains and strong light on the matching accuracy,an image restoration method used to repair the identified stains and strong light,eliminate the interference on matching,and match the light holes and threads combination gray matching methods and edge features.Finally,the subpixel coordinate fitting is carried out for the matching results.After repair,the difference between the matching accuracy of smooth hole and threaded hole and that of commercial software is less than 0.3.Figure 59;Table 14;Reference 52... |