Morphology of molten pool is significantly associated with welding quality. In this study, an active vision system was designed to form the shadow of a molten pool to get its morphology information during high-power disk laser welding of Type 304 austenitic stainless steel. The shadow was segmented from the recorded images, and four features of the shadow were defined to describe the morphology of a molten pool. The BP (Back propagation) and RBF (Radial based function) neural networks were established to model the relation between the four features and the solid weld width which is the foremost characteristic of weld quality. The effectiveness of two models were compared and analyzed at different welding speeds, and it was found that the BP model had the better results than RBF. The study focuses on predicting the solid weld width by observing the morphology of a molten pool, and provides a necessary foundation for online monitoring and control of weld width during laser welding.