Abstract: To overcome issues such as thermal deformation of welded components and interference from smoke that make it difficult to obtain accurate weld seam information, this study designs binocular vision system based weld seam spatial position information acqusition for automatic welding of non-standard large workpieces. The system includes a red-line structured laser, narrow-band red filter, and binocular vision cameras fixed together with the welding gun at the robot's effector. During welding, real-time image capture and position sensing of the seam are performed. A deep learning neural network based on the generative adversarial network (GAN) architecture is designed, and transfer learning is employed for cross-domain training. Experimental results demonstrate that the designed binocular vision system can effectively process binocular data, providing real-time seam position and depth. Without specialized image calibration, the recognition accuracy of the lateral position and height of the weld seam during the welding process can reach 1.0 mm. The proposed binocular vision method for welding robots is concise, feasible, and cost-effective.