We propose a wavelet multi-resolution analysis to localize specific features in both lateral and frontal radiographs. This analysis allows an elegant spectral investigation that leads simultaneously to image de-noising and edge extraction. It is combined with an a priori knowledge of the spine's morphology and a 3D spline curve characterization of its global shape. Actual work deals with identifying the contours of the vertebral bodies and the localization of vertebrae’s endplates. However, this information could also lead to the selection of a 3D statistical model of the spine suited for the studied deformation. Working with retro-projections of the model, we aim at creating edge models for each vertebra that will be used to geometrically match the wavelet's edges. The manual feature identification could then be replaced in the reconstruction of the 3D representation of the spine.