Research On Vehicle Parts Defect Detection Based On Deep Learning

The popularization and development of intelligent manufacturing make intelligent the production,manufacturing,testing and maintenance of automobiles.As automobiles are a common means of transportation at present,systematic testing and strict inspection should be carried out in the process of assembly,manufacturing and production of vehicles for security reasons.Therefore,the defect detection for vehicle parts is particularly important,which has developed from manual detection to image processing algorithm combined with traditional feature extraction algorithm,and to machine vision method.However,most of the methods are targeted at a specific defect or a defect in a certain position,and the identification algorithms for multiple defects are relatively rare.So,some materials still need to be re-tested manually.To solve the above problem,this paper studies four common defects in automobile assembly:drive shaft defects,shock absorber spring defects,positioning bolt defects and shock absorber defects,and then proposes an algorithm of vehicle parts defect detection based on deep learning.Firstly,the image data set of vehicle parts should be processed and enhanced.Because the background of vehicle parts' pictures is often complicated and contains many disturbing elements,such methods as erosion,dilation and median filtering are used to process the experimental pictures,while such image processing methods as contrast enhancement,scaling,rotation,vertical and horizontal mirroring are used to enhance the data.Secondly,the convolutional neural network is used to make classification.In reference of VGG16 structure model,simplifies the complex model structure and builds a small VGG network model,S-VGG.On this basis,in combination of the idea of asymmetric convolution in Inception V3 model and the idea of fusion of hidden layers,a convolutional neural network model SF-VGG with fusion layers is designed.Then the idea of densely connected network is added to the model,which expands the depth of the network and constructs the Dense SF-VGG network model.In this paper,sufficient experiments have been done on the SF-VGG network model and the Dense SF-VGG network model.It is proved by the experiments that the accuracy of the SF-VGG network model and the Dense SF-VGG network model on the test set is 98.36% and 98.73% respectively.And its comparison with Goog Le Net,VGG16 and S-VGG prove its superiority.Finally,the robustness of the model is verified by fuzzy picture test experiments and irrelevant data experiments.The accuracy of fuzzy image test experiment on the fuzzy image data set is 95.75%,and the accuracy of irrelevant data experiment on the irrelevant picture data set is 91.4%.The experimental results conclude that the model as proposed by this article is effective,accurate and feasible in vehicle parts defect detection.