A method was proposed to obtain characteristic information in a welding process by visual sensing and to predict the weld width of weldment bottom surface by using a neural network model. A workpiece made from mild steel SS400 was welded by a high power disk laser. In welding processing, the weld conditions were changed, including laser welding power, welding speed and welding route and two high speed cameras were used to capture images containing characteristic information on both top surface and side surface of weldment simultaneously. In order to get a better characteristics extraction, the colour space of a RGB image was changed into NTSC (National Television Standards Committee) colour space, then both RGB image and YIQ image were separated into their colour components, filtered to denoising and processed in space domain. The weld characteristic information was extracted, including spatter, weld pool and metal vapour and the effect of weld route on characteristic information was researched. Finally, a LMBP (Levenberg-Marquardt Back Propagation) neural network model including three layers and one hidden layer was established. The obtained characteristic information was taken as input, and the weld width of weldment bottom surface was predicted. The results show that when the welding penetration is unstable or lack of penetration, the fitting degree of LMBP neural network is greater than 0.83, the maximum training error mean is 0002 8 mm, and maximum actual error mean is 0.225 6 mm. It concludes that the prediction model has good accuracy and stability.