Weld Surface Hole Detection Based On Machine Vision

The Additive Manufacturing technology,which is known for its low cost and short cycle,has outstanding application advantages and obvious service benefits in aerospace and other manufacturing industries.However,the violent manufacturing process may cause defects in AM parts and affect the mechanical properties of the products or even failure.Among them,hole is one of the most common types of weld surface defects,and the number,size and position distribution of hole are not significantly regular,which makes it difficult to automatically detect various types of hole.In order to solve the above problem,this paper proposes a hole defect detection algorithm combining "coarse positioning + fine positioning".Coarse location method of hole defect region.Because the difference of weld seam foreground and background is large and the edge feature of hole defects are obvious,this paper uses a feature extraction algorithm based on image gradient information--HOG algorithm.Firstly,the sample data is analyzed and the image preprocessing is carried out in a targeted manner.Then the image feature vector is extracted by HOG algorithm.Finally,the support vector machine is used as the classification model to learn data features and complete the coarse positioning of the hole region.Since the collected samples are unbalanced data,the ROC curve is used to represent the performance of the model in the verification set,and AUC value is calculated to quantify the classification result of the model.The average AUC value calculated by the 6-fold cross-validation method is 91%,which is also the accuracy of the coarse positioning of the model region.Fine positioning method of hole coordinates.The defect region obtained by the HOG + SVM algorithm is taken as input,and the SIFT algorithm is used to extract and detect the hole feature in the region,and the center coordinates and radius of the fitting circle of the defect are obtained.In order to feed the detection data back to the machine tool for secondary processing,this paper also carries out pixel equivalent calibration which converts the pixel coordinates and radius obtained by the algorithm into physical coordinates and length.Under the experimental conditions,the comprehensive positioning accuracy of the algorithm reaches 87%.