Research Of Workpiece Defect Detection Based On Faster-RCNN

In today's social production process,various types of workpieces are widely used in the automotive industry,aerospace and other fields.Workpiece defect detection is an important part of ensuring the quality of workpiece production.Solving this problem can effectively improve the production efficiency of the workpiece.With the development of computer hardware and detection algorithms,the detection of workpiece defects is gradually replaced by machine vision inspection by the initial manual visual inspection.The basic idea is to obtain the position of the defect by comparing the image of the workpiece and the standard workpiece,and then to the defect.Feature extraction is performed,and finally the classifier is designed to identify the defect type.However,due to the variety of workpieces and the different types of defects,feature extraction is the main problem faced by such methods.Therefore,the current industrial detection algorithms lack certain universality.In view of this engineering problem to be solved,this paper studies the method of workpiece defect detection based on deep learning.The main work contents are as follows:(1)The sample defect expansion and preprocessing method of workpiece defect is studied.For the shortage of sample size,the sample is expanded by the combination of traditional image method and generated anti-network.The Laplacian sharpening and homomorphic filtering algorithm is used to preprocess the image,and the preprocessing effect is judged according to entropy,standard deviation,gray mean and average gradient.The processed image is manually labeled,and the VOC2007 format data set is created and completed.(2)For the complex workpiece defects,the Faster-RCNN network model is used to identify and locate the workpiece defects.The basic network models such as ZFNet,VGG-16 and ResNet-50 in the convolutional neural network are verified and analyzed by using the fabricated texture defect dataset.The network model of ResNet-50+FasterRCNN is determined.In view of the fact that the deep features in the convolutional network have a certain abstraction and the receptive field is relatively large,it is a kind of global information.After the multi-level convolution operation,many details are lost.This paper adopts a multi-level feature fusion method.Generate new features,and the experiment proves that it has stronger feature representation ability.(3)In order to verify the universality of the proposed algorithm model,the migration learning scheme is used to identify the hot-rolled strip surface defect dataset.The results show that the model still has a high detection rate on such workpiece defects.However,due to the different sizes of hot-rolled strip defects,the fixed anchor frame setting scheme leads to the model convergence speed being too slow and the positioning has a large deviation.To this end,K-means++ clustering method is used to optimize the selection of anchor frames in RPN networks,and the models before and after optimization are compared and analyzed.It is proved that the optimization method has the advantages of speeding up model convergence and improving positioning accuracy.(4)Finally,based on the Pycharm development platform,PyQt5.12 and MySQL database,the workpiece defect detection simulation platform is built,and the proposed model is tested and analyzed on the simulation platform.The results show that the accuracy of the model used in this paper can reach more than 90%,and the coincidence rate of the bounding box and the marking frame can reach more than 50%,which is in line with the needs of engineering applications.