This paper presents improvements to the laser/air-coupled noncontact ultrasonic method for detecting transverse cracks in rails. The work advances the development of a rail inspection prototype based on non-contact probes and ultrasonic guided waves for use in the field at high inspection speeds. Surface waves are used in the range of 100 kHz - 900 kHz to detect and size surface-breaking cracks located at the centre of the rail head and at the gauge-side corner of the rail head. The raw ultrasonic signals are processed through discrete wavelet de-noising to extract various features that are related to the crack size. These features are statistical parameters extracted from the wavelet coefficient vectors, and other parameters extracted from the transforms (FFT and HT) of the wavelet reconstructed signals. Based on these features, a multidimensional damage index vector is then constructed and fed to a neural network for classifying the cracks by size into three classes: pristine structure, below 10% HA reduction, and between 10 % and 20 % HA reduction. The current configuration of the system uses two different inspection strategies, one based on reflection measurements and the other based on transmission measurements. The laboratory results show that the classifier performs well with the proper selection of the defect-sensitive features and of the network design parameters. The best classification was obtained in the transmission mode, with 95.6 % total success rate among the three classes considered, and only 5 % of false negatives using four features. In the reflection mode, the best total success rate was 76.7 % using three features. The defect location, in addition to the defect size, is being added to the output vector of the classification algorithm in the prototype. A more realistic assessment of the performance of the feature selection and the classification algorithms will be made once the prototype is tested in the field.