Abstract: Quantitative recognition of defect was the difficulty in pipeline magnetic flux leakage(MFL) inspection. The relationship between pipeline defects and MFL signals were studied. The study showed that the depth and length of the defects had an approximately linear relation to the amplitude and width of MFL signals respectively. The real MFL signals were preprocessed to eliminate the effects of sensor lift-off and then were recognized quantitatively by BP neural network already trained. The recognition result error was less than 10%, and practical inspection requirement was completely fulfilled. In order to get better recognition result, the axial and radial MFL signal were fused at signal level by weighted average and adaptive weighted average methods respectively. The accuracy and reliability of quantitative recognition based on BP neural network were improved remarkably after signal fusion. The results showed that the weighted average method was better than the adaptive weighted average method.