Study on HOG-SVM detection method of weld surface defects using laser visual sensing

In order to realize automatic detection and classification of weld surface defects, an effective laser vision detection method for weld surface defects was studied. First, the weld image is collected by a laser vision sensor and preprocessed, including image segmentation, grayscale, smooth denoising and weld contour extraction. Then the feature vectors of laser stripe contour image of weld seam are extracted by means of Histogram of Oriented Gradient (HOG), the model parameters were optimized based on the five-fold cross-validation grid search method. Finally, an intelligent model of Support Vector Machine (SVM) was established to identify and classify weld surface defects. Different feature data were obtained by adjusting the weld contour extraction algorithm and HOG feature dimension, and then the identification effect of weld defects was analyzed by comparing each other. Under same experimental conditions, it is found that the recognition rate of SVM is higher than that of random forest classifier, K-nearest neighbor classifier and naive BAYES classifier, reaching 97.86%. The proposed intelligent identification method of weld surface defects based on HOG-SVM can effectively improve the classification accuracy of weld defects (porosity, sag, undercut) and non-defects.