Abstract: An algorithm of Kalman filtering compensated by neural network(NN-KF) of magneto-optical imaging is researched for detecting micro-gap weld joint. A motion model is proposed and the Kalman filter is applied to predict the weld position deviation optimally. To accomplish the optimal estimation by using Kalman filter, an accurate system model and an observation model need to be established. However, the system noise has the characteristic of non-apriority in seam tracking process. Considering the influence of the motion model error, the process noise and the measurement noise on Kalman filtering, a BP neural network was employed to amend the Kalman filtering results through compensating for the filtering error caused by model error and the uncertainty of noise statistic characteristics. Experimental results demonstrate that Kalman filter compensated by a BP neural network can effectively restrain the divergence of Kalman filtering, reduce the influence of noise, and improve the seam tracking accuracy.