Relational Data Base Management Systems (RDBMS) are currently the most popular database systems. The relational model of data is a simple and powerful model for representing real world applications. However, it lacks the expressiveness present in conceptual models. Unlike the latter, the relational model does not make an extensive use of the generalization abstraction. Therefore, it does not allow the designer to represent directly a large variety of integrity constraints. Moreover, inclusion dependencies formalizing inter-relational constraints, can not directly be represented in the relational model, due to the fact that the relation is its unique building structure. Finally, relational databases lack concepts that allow the designer to represent inheritances in a natural way. A number of approaches have been proposed to extract from relational databases specialization/generalization links. In this paper we make a comparative study of six representative previous approaches which provide methods to reverse-engineer the inheritance links hidden in a relational database. We then present a new improved and more general method combining heuristic and algorithmic techniques, allowing us to go beyond the limits of previous approaches. A common example is used to illustrate and compare all the approaches. A set of criteria is proposed to facilitate this comparison.