At present, micro flat motor manufacturers still use manual observation of motor FPC surface welding quality for classification, its detection accuracy is low, slow speed. To solve this problem, a defect classification detection method based on improved Faster R-CNN was proposed. Firstly, the last two layers of VGG16 are fused by multi-scale feature fusion network to replace the input feature graph of the regional proposal network in the original Faster R-CNN. Then, the accuracy, recall rate and score of the network are compared from three multi-scale feature fusion algorithms with different depths. The experimental results show that the average accuracy of defect classification detection of the improved two-layer multi-scale fusion feature map is 91.89%, 7.72% higher than that of the traditional model. Compared with the other two models, the improved model has the highest classification detection accuracy and precision.