Abstract: This paper utilizes the computer to assist in the weld defect detection of the in-service pipeline,in the defect feature,degree of circularity,length-width ratio,compactedness,sharpness of the nose,degree of symmetry、gray scale ratio and the position of the defect’s barycentric coordinates relative to weld center etc.seven parameters are extracted as defect feature,so that different fault defect can be classified and recognized.In the solution of the defect classification and recognition,this paper adopts the self-organized and self-adaptive-3 layers back propagation neural network,applies modified BP algorithm,and takes weld defect feature parameters as training sample of neural network.The network’s hidden nodes number,momentum coefficient,error level and step length etc.network parameters can be obtained optimum values by experimental method.Finally,weld defect classification and recognition achieved effectively.