We present the network we use to detect and recognize symbols (doors, windows...) in architectural drawings. Our network is based on Messmer's approach to exact and inexact graph matching and we propose some modifications to this method for our specific document analysis problem. First, a description language has been written in order to describe different kinds of constraints on the primitives of the symbols. Moreover, our network is not limited to mere graph isomorphism and is able to handle these new constraints. This work is still in progress, and a lot of things must be improved~: post-processings to improve the vectorized picture, network construction based on heuristics, important run time memory and load, experimental error measures...
Keywords: network, matching, constraints.