Detecting the surface and subsurface micro weld defects is the key to ensure welding quality. A weld defect detection method with magneto-optical imaging based on deep convolutional network is proposed. On the basis of Faraday magneto-optic rotation effect, the principle of magneto optical imaging is analyzed. A deep convolutional network prediction model is established to study the influence of different model structure parameters on the training results. Through analyzing the intermediate mechanism of deep convolutional neural network, the model training process is studied and the optimal parameters of convolution kernel are found automatically. Experiment results show that the optimal prediction model can be achieved by selecting the size of the first layer convolution kernel as (7×7) and using the Relu activation function. The average training accuracy of magneto-optical imaging of weld defects is 98. 61% , and the prediction accuracies of 5 weld samples with pit, crack, incomplete penetration, incomplete fusion and non-defect are 84. 38% , 98. 05% , 84. 38% , 100% and 100% , respectively, and the average prediction accuracy is 93. 36% .