Research On Visual Recognition Technology Of Rivet Appearance Defects Based On Deep Learning

Rivet as a kind of fastener,often used in aerospace,aircraft assembly and other important fields,so it is particularly important to ensure the good quality of rivet.At present,the flaw detection method based on machine vision requires artificial design of feature extractors to extract the feature information of defects.The quality of the feature extractors will directly affect the detection effect,and different feature extractors with poor adaptability and low robustness should be designed for different defects.In order to reduce the influence of human factors and improve the detection effect,the deep learning algorithm is introduced into the rivet appearance defect recognition in this thesis.The specific research contents are as follows:Firstly,aiming at the defects of rivet appearance,a visual acquisition system of rivet appearance defects is designed,which mainly collects the defect images on rivet caps.An appropriate image acquisition platform is built,analysis and selects the hardware used in the platform,mainly including the selection of industrial camera,lens,light source,and lighting way.Secondly,pre-process the collected images that the rivet surface defect,mainly including image filtering and image enhancement.Among them,image filtering mainly includes mean filtering,median filtering and bilateral filtering.After experimental comparison,the bilateral filtering method is selected to de-noise the image finally.This method not only de-noises but also protects the edge information of the defect part;Image enhancement has two methods include based on Retinex and morphology.The experiment compares the advantages and disadvantages of the two methods,and the SSR method based on Retinex is selected to eliminate the influence of uneven illumination,in order to achieve the purpose of enhancing the defective part.Thirdly,in this thesis,aiming at the problem of incomplete feature extraction of rivet defects from conventional models,a feature fusion surface defect detection method based on CNN is proposed.This method extracts feature information in parallel with the convolution part of Res Net and Dense Net and integrates the two features by adding the auxiliary residual network.It replaces the convolution add pooling feature extraction method in CNN,and enriches the feature information.The experimental results show that the accuracy of the detection method proposed in this thesis reaches 98.75%,which is about 10% higher than the conventional CNN network,and the effectiveness and feasibility of the algorithm are preliminarily verified.Fourthly,aiming at the problem that some small and inconspicuous defects on the rivet surface are easy to be lost and different kinds of defects are confused,this thesis proposes a model fusion method about rivet defect detection based on MLP.This method mainly uses several different models to extract the feature information of defects,combines the feature information of different levels to improve the recognition of the feature information,then sends all the feature information into MLP for fusion and extraction of the main features,and then uses softmax function for classification.The data set of rivet surface defect was used to classify and identify Alex Net,VGGNet,Goog Le Net,Res Net and Dense Net,and the network of VGGNet,Goog Le Net and Dense Net with good performance was selected for fusion.Experimental results show that the detection accuracy can reach 99.06% after fusion,which verifies the effectiveness of the algorithm.