This article deals with the analysis of pages of French parish registers from 16th to 18th century. These documents are structured in paragraphs called acts. Each act contains valuable demographic information that can be useful to genealogists willing to find information about their ancestors. The first step toward parish register analysis consists of delimiting each act. But these documents are so poorly-structured that the visual separation between the acts is not always clearly visible. One of the main visual indication of separation is the signature of the priest at the end of each act. In this work, we propose to train and compare several u-shaped neural networks for signature detection. We also propose a rule-based system for segmentation into acts and evaluate the impact of signature detection at act level.