Research On Image-Based Evaluation Method Of Drop-Weight Tear Fracture Of Steel

To assess the Drop-Weight Tear Fracture of petroleum pipe steel,the current methods are mainly measured and calculated by measuring tools such as vernier calipers.There are high requirements for staff experience,manual determination of subjective factors,and irregular shape.And low efficiency.Aiming at the problems existing in the evaluation of steel fractures above,this paper starts from two aspects: the traditional image segmentation method and the machine learning direction segmentation method.1)Research on steel fracture morphology,starting from traditional image processing methods,using edge-based segmentation method,area-based segmentation method and background-based segmentation method to conduct experimental research on steel fracture image in many aspects,aiming at steel fracture morphology The more complicated samples use interactive image segmentation method to select and calculate the brittle area in the steel fracture image.2)A method for semantic segmentation of steel fracture images based on a full convolutional neural network is adopted.First,the fragile region data set is labeled for the collected steel fracture image,and then the FCN network model is trained using the labeled data set.The trained model is used to predict the fracture image of the steel.By comparing the predicted result image and the labeled image,it is found that the full convolutional neural network model can identify the location of the brittle region in the fracture surface,but there is brittleness to the fracture surface of the sample.The boundary of the area is not fine enough,and it is not sensitive enough to the details in the image.3)A method for semantic segmentation of steel fracture images using a codec model with hole convolution is adopted.In order to train on the tensorflow machine learning platform more effectively,the prepared data set is converted into Tfrecord format.After the Deep Lab V3+network model is trained,the model can effectively segment the brittle region in the fracture of the sample.In order to visually predict the fractured image of steel,the model segmentation method of Mask RCNN is used to train the model.The trained model is evaluated and found that the method has higher detection efficiency and segmentation effect.4)Finally,based on the calculation formula for the evaluation of steel fractures in the standard of GB/T 8363 "Ferrite steel drop weight tear test method",combining the characteristics of the original fracture image,marked image and identification image of the steel fracture,the calculation method of the fracture area percentage of the steel sample is improved based on the pixel level.After analysis and comparison through experiments,it proves that the method used in the experiment is effective.The method for identifying the brittle area of steelfractures based on image processing has higher stability,accuracy and segmentation effect.