Handwritten digits recognition has been widely studied because of its potential application in automatic sorting of mail pieces. In this paper, we focus on on-line isolated digits with unknown scriptor. TCSF/LER has developed an intermediate approach between classical methods, based on extracting small sets of parameters, and pure neural methods, in which the network is fed with raw image data. The proposed method combines image processing and connectionnist recognition. A vector of 90 parameters consisting in profile curves, measures of density and morphological information is computed from the digit image. Then a multilayer perceptron trained by backpropagation is used to classify. The method has been evaluated on a huge database of real zip codes, provided by the SRTP Nantes. The database includes around 20000 digits for learning and 12000 digits for testing. The isolated digits come from prefilled or free envelopes. Each digit has two labels provided by two operators: the first one sees the whole address block and the second one is restricted to seeing only the segmented digit. In this paper, we describe our approach and we give many experimental results: recognition rates on prefilled envelopes, on free envelopes, on digits confirmed by the second operator, etc.