Weld recognition based on key point detection method

In order to ensure the quality of automatic welding and improve the accuracy and adaptability of weld identification, a key point detection method for weld feature extraction is proposed. A weld feature extraction network was designed based on the convolutional neural network. The network extracted weld feature by convolution and pool operation. The feature map from deep layer is sampled up, and then the feature map from deep layer and shallow layer are fused to improve the accuracy of weld feature extraction. The feature point position of weld seam is predicted by the thermal image of weld seam, and the recognition and location of many kinds of groove weld seam are realized, and eliminating the need for non-maximum suppression algorithm, which improves the feature extraction speed. The network model is trained by collecting different weld feature images. The experimental results show that the root mean square error of weld feature point location is 0.187 mm. The network model designed in this research has high detection accuracy in the weld feature point recognition task, and has strong adaptability and generalization, and meets the requirements of automatic welding.