Method for performing cross-hand palmprint recognition
The method for performing cross-hand palmprint recognition uses a semi-Siamese neural network to compare a first hand (e.g., a right hand) anchor embedding, which is obtained through feature aggregation of first hand images, to a second hand (e.g., a left hand) query embedding. The semi-Siamese neural network's branches share a feature-extraction backbone and employ distinct final-layer weights to allow hand-specific calibration while maintaining a common embedding space. A hybrid similarity is computed, which includes both an element-wise absolute-difference (EAD) vector and a Euclidean-distance scalar. These are concatenated and input to a classifier to produce a match score or decision regarding authentication.
1 . A method for performing cross-hand palmprint recognition, comprising:
storing a plurality of palmprint images of a first hand of an authorized user in a non-transitory computer readable memory;
obtaining a palmprint image of a second hand of an unidentified user and storing the palmprint image of the second hand in the non-transitory computer readable memory;
performing feature extraction on each of the plurality of palmprint images of the first hand and the palmprint image of the second hand using a semi-Siamese neural network having a shared feature-extraction backbone and distinct final-layer weights for the plurality of palmprint images of the first hand and the palmprint image of the second hand;
extracting a fixed length feature vector from each of the plurality of palmprint images of the first hand and the palmprint image of the second hand;
forming an anchor embedding through feature aggregation by averaging the fixed length feature vectors extracted from a number of the plurality of palmprint images of the first hand to obtain a stable first hand anchor embedding;
computing a hybrid similarity between the stable first hand anchor embedding and the one of the feature vectors corresponding to the palmprint image of the second hand, wherein the hybrid similarity comprises an element-wise absolute-difference vector and a Euclidean-distance scalar;
concatenating the element-wise absolute-difference vector and the Euclidean-distance scalar to produce a combined similarity vector;
classifying the combined similarity vector to output a match probability between the palmprint image of the second hand of the unidentified user corresponding to the stable first hand anchor embedding associated with the first hand of the authorized user; and
determining that the unidentified user is the authorized user when the match probability is above a threshold probability wherein the feature-extraction backbone comprises a visual geometry group palmprint-tuned convolutional neural network and the distinct final-layer weights are confined to a final dense layer of each of a first hand branch and a second hand branch of the semi-Siamese neural network.
2 . The method for performing cross-hand palmprint recognition as recited in claim 1 , wherein the number of the plurality of palmprint images of the first hand used to form the anchor embedding comprises between two and five of the plurality of palmprint images of the first hand.
3 . The method for performing cross-hand palmprint recognition as recited in claim 1 , wherein the number of the plurality of palmprint images of the first hand used to form the anchor embedding comprises four of the plurality of palmprint images of the first hand.
4 . The method for performing cross-hand palmprint recognition as recited in claim 1 , further comprising normalization of the element-wise absolute-difference vector and the Euclidean-distance scalar prior to the concatenation thereof.
5 . The method for performing cross-hand palmprint recognition as recited in claim 1 , further comprising scaling of the element-wise absolute-difference vector and the Euclidean-distance scalar prior to the concatenation thereof.
6 . A method for performing cross-hand palmprint recognition, comprising:
storing a plurality of palmprint images of a first hand of an authorized user in a non-transitory computer readable memory;
obtaining a palmprint image of a second hand of an unidentified user and storing the palmprint of the second hand in the non-transitory computer readable memory;
performing feature extraction on each of the plurality of palmprint images of the first hand and the palmprint image of the second hand using a semi-Siamese neural network having a shared feature-extraction backbone and distinct final-layer weights for the plurality of palmprint images of the first hand and the palmprint image of the second hand;
extracting a fixed length feature vector from each of the plurality of palmprint images of the first hand and the palmprint image of the second hand;
forming an anchor embedding through feature aggregation by averaging the fixed length feature vectors extracted from the plurality of palmprint images of the first hand to obtain a stable first hand anchor embedding;
computing a hybrid similarity between the stable first hand anchor embedding and the one of the feature vectors corresponding to the palmprint image of the second hand, wherein the hybrid similarity comprises an element-wise absolute-difference vector and a Euclidean-distance scalar;
concatenating the element-wise absolute-difference vector and the Euclidean-distance scalar to produce a combined similarity vector;
classifying the combined similarity vector to output a binary decision indicating whether the palmprint image of the second hand of the unidentified user corresponds to the stable first hand anchor embedding associated with the first hand of the authorized user; and
determining if the unidentified user is the authorized user based on the binary decision;
wherein the feature-extraction backbone comprises a visual geometry group palmprint-tuned convolutional neural network and the distinct final-layer weights are confined to a final dense layer of each of a first hand branch and a second hand branch of the semi-Siamese neural network.
7 . The method for performing cross-hand palmprint recognition as recited in claim 6 , wherein the binary decision is based on a decision threshold selected to satisfy a target false-acceptance rate and/or a false-rejection rate.
8 . The method for performing cross-hand palmprint recognition as recited in claim 6 , wherein the binary decision is based on a decision threshold selected to minimize an equal error rate.
9 . The method for performing cross-hand palmprint recognition as recited in claim 6 , wherein the number of the plurality of palmprint images of the first hand used to form the anchor embedding comprises between two and five of the plurality of palmprint images of the first hand.
10 . The method for performing cross-hand palmprint recognition as recited in claim 6 , wherein the number of the plurality of palmprint images of the first hand used to form the anchor embedding comprises four of the plurality of palmprint images of the first hand.
11 . The method for performing cross-hand palmprint recognition as recited in claim 6 , further comprising normalization of the element-wise absolute-difference vector and the Euclidean-distance scalar prior to the concatenation thereof.
12 . The method for performing cross-hand palmprint recognition as recited in claim 6 , further comprising scaling of the element-wise absolute-difference vector and the Euclidean-distance scalar prior to the concatenation thereof.