Current machine-learning techniques for Layout Analysis require extensive annotated data for training. Although there are proposals to reduce the amount of annotated data, they are often limited to simple criteria for selecting regions to annotate. This work conducts a comparative study of four criteria for selecting patches to be annotated. The main objective is to analyze their influence on performance and the amount of data needed. The proposals are evaluated on four datasets, showing an average improvement of 6.7% over a random extraction method and 12% over the sequential strategy commonly used in the state of the art.