Aiming at the low accuracy of crack depth prediction in metal surface utilizing the empirical fitting method based on specific ultrasonic signal parameter and object characteristic, based on the divergence analysis and least parameter set principle, as well as training BP neural network, a surface acoustic wave (SAW) quantitative characterization technique of metal surface crack depth is proposed. This technique simulates the laser exciting SAW process with finite element method and extracts the characteristics of peak and mean values of the reflected and transmitted SAW signals caused by the surface crack, which is used to train BP neural network and predict the crack depth. The quantitative characterization of 20 groups of the opening cracks with crack depth of 01~20 mm on stainless steel specimen surface was realized. The simulation results show that the relative errors of the predicted crack depth are within 3%. Compared with the prediction result of empirical fitting curve, the accuracy is improved by more than 60%. Experiments adopted a 5 MHz SAW transducer to acquire 20 reflected SAW signals of two preprocessed cracks of the stainless steel specimens at surface depth of 10 and 15 mm, respectively. The relative errors of the crack depth predicted by BP neural network are within 01%, which verifies the feasibility and accuracy of the proposed quantitative characterization technique.