Extraction and choice of features are critical to improving the recognition rate of off-line handwritten digits. Principal curves are nonlinear generalizations of principal components analysis. They are smooth self-consistent curves that pass through the “middle” of the distribution. They preferably reflect the structural features of the data. During digit feature selection, firstly principal curves are used to extract the structural features of training data; Secondly the classification features used for digits coarse classification and precise classification are chosen by analyzing the structural features of principal curves in detail; Finally coarse classification and precise classification are separately carried out in handwritten digits recognition. The Concordia University CENPARMI handwritten digit database is used in the experiment. The result of the experiment shows that these features have good discriminating power of similar digits. The proposed method can effectively improve the recognition rate of off-line handwritten digits and provide a new approach to the research for off-line handwritten digits recognition.