Abstract: In order to realize the complete three-dimensional sensing of weld pool surface morphology, a stereo vision sensing system with biprism and single camera was constructed. Aiming at the difficulty of stereo matching caused by the lack of texture in weld pool map, a globally optimized variational stereo matching algorithm was introduced. By establishing the feasibility function of energy function containing gray difference data item and spatial continuity constraint item, the dense disparity figure of weld pool surface with rich details was obtained through iterative. The results of stereo matching and three-dimensional reconstruction of the self-made non-standard concave shape show that the width error is less than 3.16% and the depth error is less than 4.82%. Based on this algorithm, the dense disparity map of weld pool surface is calculated and reconstructed under the conditions of bead on plate and V-groove butt welding with different penetration states.