The invention discloses a method for extracting a key pattern from an image file. The method is characterized by comprising the following steps of: acquiring an original image file and determining a template S; determining a neighborhood direction angle reference characteristic X', a gravity centre included angle reference characteristic Y' and a gravity centre distance reference characteristic Z' in the template S and establishing a reference table; acquiring the neighborhood direction angle characteristic X of each pattern pixel in a processed image M and finding out a matching pattern pixel D matched with the neighborhood direction angle of the neighborhood direction angle reference characteristic X'; and citing the gravity centre included angle reference characteristic Y' and the gravity centre distance reference characteristic Z' and performing reduction voting on a virtual key pattern in the processed image M to position a real key pattern. The method has the advantages of breaking through the limitation that shape matching needs complete profile information, simplifying the calculation of the local characteristic, effectively extracting handwritten keywords with local deformation, partial rotation and scaling in the file image and fulfilling the aim of identifying the identity of a writer by modeling the writer by a signature identification method.