A method is proposed to identify and localize semantic features like anatomical characteristics or decorations on digital artefacts or fragments, even if the features are partially damaged or incomplete. This technique is based on a novel generalization of the Hough transform. Its major advantages are the relative robustness to noise and the recognition power also in the case of partial features. Our experiments on digital models of real artefacts show the potential of the method, which can work on both 3D meshes and point clouds.