Steel Pipe Weld Defect Type Detection

In recent years,in the field of industrial welding,welding technology has been continuously developed,steel pipes and other weldments have been put into production applications.Weld defects refer to defects that occur in the welding process due to technical factors or external environmental factors in actual operation.Weld defects are great hazard,and in any area,the weldments must be guaranteed to be free of defects.Under the premise of ensuring the identification of defects,the defects need to be repaired.The premise of the repair is to clearly know the type of the defects.This paper improves the system function based on the automatic defect identification system of a certain steel pipe plant,and give two kinds of classification algorithm which can automaticly find the type.If use artificial methods,subjective judgments may be flawed,the efficiency is very low.Based on the X-ray imaging,a set of automatic inspection system for weld defects was developed.Firstly,the Imaging method,working principle and detection algorithm of this system were briefly introduced.Since the operating objects of this experiment are all static images,the image imaging characteristics of the DR system need to be summarized.Secondly,the process of extracting shape features and gray features is given.By understanding the basic functions and principles of the system,various types of defect images detected by the system are collected as sample sets.A series of operations such as dilating,clustering,and contouring are performed on the original image.The binary image is used to clearly find the position of each defect and the set of points it contains.According to the formation reasons and shape characteristics of various types of defects,we give the criterion of defect identification,some morphological feature parameters and gray feature parameters are selected as the feature description vector of the sample.The second idea of extracting features is based on the original HOG feature idea.Firstly,aiming at the shortcoming of the traditional HOG feature,the cell division method was improved to a circular division method,and then the RGT transformation was used to make statistics on the gradient histogram,and finally the relevant parameters were determined.In order to solve the shortcomings of the traditional HOG feature scale variability,the object can be processed using a scale normalization method,Finally,the LSSVM model is used to train and identify the above two features respectively.Because the feature dimension of the improved HOG feature extraction is high,it is necessary to reduce the length of the sample before using the improved HOG feature with rotation invariance to perform sample training,here we use PCA method to reduce it.Compare the results of the two methods to get the final conclusion.The final result achieved the desired effect.