In practice, traditional metal electromagnetic detection technology is insufficient to detect sheet metal internal defects. In order to solve this problem, a method of recognizing surface and internal defects based on deep learning and electromagnetic detection is proposed in this paper. In our experiment, nine types of surface and internal air-gap defect models with different depths, positions and quantities are built. The feasibility of multi-frequency detection is verified. The relation between excitation voltage and induction voltage detected by array electromagnetic sensor is concerned. Data balance acquisition method is introduced to expand data set before data preprocessing. In order to get the characteristics of measurement data, DNN and CNN are constructed. The appropriate network with suitable parameters is chosen according to the recognition results. The experimental results show that 9 kinds of air-gap defects with different quantities, positions and depths can be recognized by electromagnetic detection applied with DNN or CNN. The recognition accuracy is over 90%. Compared with CNN, DNN is faster and more efficient to classify.