A multi-scale wavelet edge extraction algorithm and Principal Component Analysis-Back Propagation(PCA-BP) neural network classification model were proposed based on magneto-optical imaging to detect the welded defects such as sags, insufficient fusion on subsurface and welding misalignment. The visualization of detection and the classification of welded defects on the surface and subsurface of weldments were explored. Firstly, the weldments were magnetized by using an excitation magnetic field. Meanwhile, a magneto optical (MO) sensor based on the principle of Faraday magneto effect was used to acquire the MO images of weldments with welded defects. Then, a defect edge extraction algorithm with a better anti-noise property was investigated based on wavelet modulus maxima multi-scale information fusion theory to process MO images suffered from serious noises, low contrast and complex background. Finally, the PCA was taken to preprocess the column grey variables of MO images and 256 feature points of column variable of MO images which could characterize grey variable by 95% were obtained. Furthermore, these feature points were regarded as inputs of a three-layer BP neural network model to classify the welded defects. Experiment results show that the proposed method can be applied to detection of welded defects as mentioned above, and the accuracy of PCA-BP classification model has reached to 90.80%.