Abstract: This paper presents a method based on RBF neural network to improve the quantification ability of magnetic flux leakage detection for crack defects of oil tank floor. Firstly, the finite element simulation was used to calculate the leakage magnetic flux signals of different lengths, widths, depths and inclination. The distribution law of magnetic flux leakage signals was analyzed and the amplitude and the width of the magnetic anomaly were extracted as the magnetic signal characteristics. The relationship between characteristic quantity of the magnetic signal and defect size was discussed and the sample set is established. Secondly, it was established that a quantitative model combining RBF neural network and simulated annealing algorithm. The RBF neural network was trained by the sample set of defects to predict the defect size and inclination. The results show that characteristic quantities of the magnetic anomaly vary with the size and angle of the defect. The complex network is established by RBF neural network, and the simulated annealing algorithm can accurately quantify the defect. The average correct rate of the defect quantization is about 99.14% in the sample set and 97.33% outside the sample set. Therefore, the method based on RBF neural network combined with simulated annealing algorithm can be applied to magnetic flux leakage detection to quantify the defects of bottom of oil tank, and provide theoretical basis for the safety assessment of oil tank.