Deformable Models for Hand Shape Matching

By Nico Duta and Anil Jain

Motivation:

We propose to approach the practical problem of person verification using the powerful tools of deformable shape analysis. This is motivated by the limited ability of the hand shape acquisition system to implicitly register different hand images using the rigid pegs on the hand scanner platen. If the user has not been properly trained or if he/she does not cooperate properly, then the resulting images are not aligned and the system's verification performance degrades . Therefore, it is necessary to align the acquired hand shapes before extracting the feature vector used for verification.

Method Used:

Given a pair of top views of hand images acquired by a hand scanner we use the following hand shape matching paradigm:
  • Peg removal :
    A mask containing the known positions of the five pegs is used to replace the pegs with a color that closely matches the background.
  • Contour extraction :
    An adaptive thresholding is applied to each image and a contour following algorithm is used to compute the shape of the hand.
  • Finger extraction and alignment :
    The five pairs of corresponding fingers are extracted from each contour and aligned separately with respect to the rigid transformations group. We chose to align pairs of fingers as opposed to the entire hand because of the following reasons: (i) a human hand is an articulated object and the motion of one finger cannot be described by a linear transformation, but rather by a set of local rigid transformations and small deformations, (ii) Computationally, it is faster to detect and align individual fingers than an entire hand.
  • Pairwise distance computation :
    Each alignment in Step 3 produces a set of point correspondences. The Mean Alignment Error (MAE) between the two hand shapes is defined as the average distance between the corresponding points.
  • Verification :
    The pair of hand shapes are said to belong to the same hand if their MAE is smaller than a threshold T. Usually, the Neymann-Pearson rule that minimizes the False Reject Rate (FRR) for a fixed False Accept Rate (FAR) is employed to compute T.


    • The images below show the alignment of two hand images obtained from the same user: