[en] Stress corrosion cracks are a huge potential security problem for steam turbine discs, and the quantitative detection of the initial cracks is significant for estimating the safety factor and the service life of the turbine discs. In this paper, the initial corrosion cracks are detected using the ultrasonic phased array, and a binary particle swarm optimization-radial basis function neural network (BPSO-RBFNN) model is proposed for the quantitative evaluation of small cracks in the early corrosion stage. The echoes reflected from the small cracks with different depths and orientation are acquired using the ultrasonic phased array. From these echo signals, some related features in the time domain, frequency domain and about nonlinearity are extracted. A quantitative analysis model based on the BPSO algorithm and the RBFNN is investigated for the feature optimization and the quantitative evaluation of the crack depth and orientation. The results showed that this method was effective in this quantitative evaluation work. (paper)