Research On Defect Classification And Detection Algorithm Based On Convolutional Neural Network

The surface defects of the product will not only affect it's aesthetics and usability,but also cause serious casualties,so the recognition of product defects is of great significance.At present,the defect recognition in the industrial field is still dominated by manual recognition and machine vision recognition based on image processing.The manual recognition method not only has a high work intensity,but also has a high rate of missed detection.The features extracted by the machine recognition method based on image processing lack robustness,which makes it unable to deal with complex and changeable defects.How to quickly and accurately recognize surface defects has become one of the research focuses and hot spots.With the rapid development of deep learning,convolutional neural network has shown great advantages over traditional methods in the field of vision.And with the development of modern imaging technology,the quality of the images gets better and better,so solving the problem of defect recognition based on the convolutional neural network has become a very promising method.In this paper,by analyzing the domestic and foreign research status of defect recognition,using aluminum defects as the data set,to study the classification and detection of aluminum defects based on deep learning convolutional neural network.Firstly,the basic principle of convolutional neural network is analyzed.And in accordance with the hierarchical structure of convolutional neural network,the main modules of convolutional neural network are deeply studied,including convolutional layer,activation function,pooling layer,fully connected layer and loss function.In addition,the training process of the neural network is analyzed.The idea of transfer learning is also introduced to solve the problem of insufficient data volume.Then,for the respective advantages and disadvantages of the mainstream Inceptionv3 and Densenet169 network,a multi-model fusion classification network is proposed.This multi-model fusion network is applied to the classification and identification of aluminum defects.Through experiments,compared with the two mainstream networks,this multi-model fusion network has obvious improvement on prediction accuracy,recall and precision of aluminum defects,providing better support for the automatic classification of aluminum defects.In addition,a visual study of the convolution kernel,feature map,and class activation of the aluminum defect classification network is conducted to further verify the effectiveness of the aluminum defect classification.Finally,the object detection algorithm based on convolutional neural network is studied,and the advantages and disadvantages of each algorithm are analyzed and compared.Comprehensively considering the detection speed and detection accuracy,the SSD network is selected as the basic network for defect detection.And for the problem that the default box setting of the SSD network lacks specificity for different data,the setting of the default box is optimized by using the k-means clustering algorithm.The optimized SSD network improves the speed and accuracy of defect detection.