Surface Defect Detection Based On Deep Learning

As the basic components of mechanical equipment,mechanical parts such as gears and screws are widely used in all kinds of mechanical equipment.When the equipment is running,if the parts themselves are defective,it will easily lead to the damage of the parts,which will cause the whole mechanical equipment tofail or even the whole equipment to be scrapped,and even endanger the life safety of the operators in serious cases.Therefore,the detection of surface quality of parts is particularly important in the process of processing and production of parts.The main problems of surface defect detection of parts focus on how to improve the discrimination ability of defect types,the accuracy of detection results,the discrimination accuracy of part edges and the accuracy of defect parameter calculation.In this paper,convolutional neural network and traditional digital image processing methods are combined to detect and calculate parts defects.Taking convolutional neural network as the main body,it can accurately identify defects and establish the approximate range of defects after training,and then calculate the defect parameters separately by using digital image processing technology for this area.The main contents include: the collection of parts images,the production of data sets,the establishment of network models,the training and verification of the network.Digital image processing method is added to the verified network model to calculate the defect parameters.Sub-pixel detection method is used to improve the calculation accuracy of defect parameters.The main contents of this paper are as follows:(1)According to the existing defect images and actual needs,the Machine Part,a part surface defect data set,is made.(2)According to the characteristics of defects in the image,the structure of convolutional neural network is changed to make its trained model more suitable for the detection environment.According to the number of data sets and defect types,the data sets are divided,and the divided training set Machine Part1 is used to train the network.(3)The traditional digital image processing method is used to calculate the defect parameters of the defect areas detected by convolutional neural network,and the sub-pixel detection method is used to improve the detection accuracy.It is verified that the m AP value of the network detection can reach 98.1%,and the detection accuracy of defect parameters can reach microns.