Principal curves are a feature extraction met hod based on the nonlinear transformation. Meanwhile, they are smooth self-consistent curves th at pass through the ″middle″ of the distribution and satisfy the ″self coincidence″. Thus, structural features of t he data can be extracted. Based on the soft K-segments algorithm for principal c urves, the skeletonization extraction of the fingerprint image is not smo oth enough, which often appears small circle and short branches. To solve this proplem, th e soft K -segments algorithm for principal curves and the specialties of fingerprint are analyzed. A new evaluation function is also proposed. And an improved soft K segmen ts algorithm for principal curves is put forward. Compared with those of the original alg o rithms, the smoothness and the accuracy of the proposed algorithm can be illustrated by experiments.