In order to solve the problem of defect identification in magnetic flux leakage testing, DKSVD dictionary learning method is introduced. By means of OMP algorithm and SVD algorithm to optimize the dictionary and the sparse coefficient, the optimal dictionary is constructed, and then the constructed dictionary atoms are combined to represent the data of the test set. The experimental results show the feasibility of the dictionary learning method in magnetic flux leakage signal recognition. And compared with the SVM algorithm, the DKSVD algorithm achieves better results.