Research On Surface Defect Detection Technology Of Seamless Steel Tube Based On Machine Learning

The seamless steel tubes surface defects is an important indicator to measure the quality which will play a decisive role in the performance of seamless steel tubes with more and more application demands.This thesis studies the surface defect detection technology of seamless steel tubes based on machine learning to identify the common surface defects of seven seamless steel tubes,by separately using support vector machine method and Tensorflow deep learning framework to identify seamless steel tubes surface defects and transplant the deep learning model to Android intelligent terminal.The main work of this thesis is as follows:1.Construct a database of seamless steel surface defects,consisting of seven types of pictures:pit,chafe,scratch,pockmark,warp,crack,normal.A total of 2156 sheets in seven kinds of pictures.2.Seamless steel tubes surface defects identification is realized based on SURF_BOW features and SVM.The local features of defects were extracted by SURF algorithm.The K-means algorithm was used to cluster the features.The BOW model was constructed to obtain the feature dictionary.Then the 1-a-1 SVM was trained to obtain the surface defect identification model of seamless steel tube.For the test sample,the average recognition accuracy of the method is 79.8%,and the single image test duration is 1.43s.3.The seamless steel tubes surface defects identification is realized based on deep AlexNet model by using migration learning technology.?Re-training the last three fully connected layers with the fine-tuning technique to obtain the recognition model.The average recognition accuracy of the model is 90.2%,and the single image test duration is 2.12s.?Combining the features extracted by the all-connected layer,and then re-training the fully connected layer with the fine-tuning technique to obtain the recognition model.The average recognition accuracy of the model is 91.2%,and the single image test duration is 2.44s.4.The practicality of identifying the seamless steel tubes surface defects is realized.Transplanting the feature fusion fine-tuning model to the Android operating system to realize off-line and on-line detection of seamless steel tubes surface defects.The test results show that in the offline detection mode,the average recognition accuracy is 91.2%,and the single image test duration is 5.84s.In the online detection mode,the surface image of the seamless steel tubes can be basically collected,cut and scanned to identify the defect,and the average recognition accuracy is 87.5%.