Weld Seam Detection Method with Rotational Region Proposal Network

Abstract

The traditional weld seam detection methods in passive vision are usually realized by detecting edges or textures of the weld seam. Since illumination conditions, weld seam types and backgrounds vary with tasks, these methods only work with specific weld types or environment. This paper raises a weld seam detection method that replaces the four main steps in traditional weld seam detection (pretreatment, thresholding, seam detection, seam fitting) with an end-to end neural network system. The method eliminates the ambiguity of original horizontal bounding box by adding an inclination parameter to the region proposal network (RPN). Compared with other methods in passive vision, our method is appropriate for accurate and fast detection of various types of weld seams in complex environment and meets the requirements for online seam detection of industrial robot.