A weld seam feature real-time extraction method of three typical welds based on target detection

Abstract

To realize strong robust, widely adaptable and highly precise weld seam feature detection, an extraction method based on the improved target detection model CenterNet is proposed. In this target detection method, the weld feature points on the laser strip are taken as the center points of the bounding box. This eliminates the need for an initial positioning frame. When dealing with multiple welds, an independent classifier is used to predict the weld type, which can avoid false detection. In post-processing, this result is used to determine the number of feature points to filter out suspicious targets. The method was tested with the welding image dataset of three typical welds. The classification efficiency of the three welds can reach 99.359 %, with an average extraction error of 1.754 pixel and an average processing time of 32.186 ms. The proposed method exhibits high performance in dealing with common welding images with time-varying noise.