Abstract : This thesis focuses on the study, conception, development and testing of a recognition system for two-dimensional handwritten structures. The proposed system is based on a global architecture that considers the problem of recognition as simultaneous optimization of segmentation, symbol recognition, and interpretation. In this framework, we have first designed a system to recognize handwritten mathematical expression. All the three problems of segmentation, recognition and interpretation are not straightforward. Segmentation is complex because of the large freedom for composing an expression, since delayed multi-stroke symbols are considered. Recognition has to face a large number of classes, and to deal with the problem of unknown pattern, and interpretation suffers for the fuzzy nature of spatial relationships. We have defined a solution which minimizes a global cost function where recognition costs and structural costs are combined and a large exploration of the space of solutions is proposed. The results are very promising and competitive compared to those of the literature. We have finally shown the generality of our approach in adapting the system to the recognition of another 2D language, which is used to design handwritten flowcharts.