The objective of this thesis is to develop a sketching-based CAD prototype for 3D shape modeling which can effectively and efficiently convert an input sketch into a 3D geometric model. In this thesis, the work concentrates on general, planar objects reconstruction from a single view line drawing. It addresses some key problems in the optimization-based 3D reconstruction. Firstly, the basic issues in formulating the compliance function are explored under the framework of support vector machine regression (SVR). Secondly, automatic relevance determination, Bayesian framework and SVR are coupled together to form a regularity selection method. An optimal regularity set is established from the commonly used regularities. Thirdly, a hybrid intelligent 3D reconstruction system is built to handle various line drawings. Line drawings are classified into 4 classes using the technique of support vector machine multi-classification. Effective reconstruction is performed by the “reconstruction experts” which are tuned to deal with each class specifically. Finally, the issues in the topological analysis of a line drawing are investigated, especially in distinguishing manifolds from non-manifolds. The experimental results confirm that the proposed methods can reliably resolve the 3D reconstruction issues. The contribution of this research is to improve the efficiency and effectiveness of 3D reconstruction with the proposed methods by applying advanced artificial intelligence techniques.