IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL.

70, 2021 5009812

Few-Shot Learning for Palmprint Recognition via

Meta-Siamese Network
Huikai Shao , Graduate Student Member, IEEE, Dexing Zhong , Member, IEEE, Xuefeng Du ,
Shaoyi Du , Member, IEEE, and Raymond N. J. Veldhuis , Senior Member, IEEE

Abstract— Palmprint is one of the discriminant biometric widely applied in daily life, such as face recognition [2] and
modalities of humans. Recently, deep learning-based palmprint fingerprint recognition [3]. As one of the unique technolo-
recognition algorithms have improved the accuracy and robust- gies of biometrics, palmprint recognition has received much
ness of recognition results to a new level. Most of them require
a large amount of labeled training samples to guarantee satis- research attention recently [4], [5]. Generally, researchers
factory performance. However, getting enough labeled data is applied signal processing methods to analyze the patterns of
difficult due to time consumption and privacy issues. Therefore, palmprint for personal authentication [6]. So far, local tex-
in this article, a novel meta-Siamese network (MSN) is pro- ture [7] and principal lines [8] have been exploited for feature
posed to exploit few-shot learning for small-sample palmprint representation. They are time-invariant with large interclass
recognition. During each episode-based training iteration, a few
images are selected as sample and query sets to simulate the variance and low intraclass variance. Therefore, promising
support and testing sets in the test set. Specifically, the model recognition results have been achieved with high universality,
is trained episodically with a flexible framework to learn both stability, and uniqueness.
the feature embedding and deep similarity metric function. Typical procedure of palmprint recognition consists of
In addition, two distance-based losses are introduced to assist image acquisition, preprocessing, feature extraction, and
the optimization. After training, the model can learn the ability
to get similarity scores between two images for few-shot testing. matching [9], [10]. Palmprint image acquisitions are usually
Adequate experiments conducted on several constrained and employed by optical cameras. Preprocessing is mainly adopted
unconstrained benchmark palmprint databases show that MSN to implement noise reduction and region of interest (ROI)
can obtain competitive improvements compared with baseline extraction. Then, several categories of feature extraction and
methods, where the best accuracy can be up to 100%. matching methods are proposed to separate different identities,
Index Terms— Biometrics, few-shot learning, information secu- e.g. encoding-based methods, structure-based methods, statis-
rity, meta-learning, palmprint recognition. tics methods, and subspace methods [9]. So far, deep learning
techniques have emerged as effective tools for automatic
I. I NTRODUCTION
visual understanding and obtained the state of the arts in

B IOMETRICS is an effective technology using human’s

physiological or behavioral characteristics for authenti-
cation [1]. Recently, several biometric modalities have been
many computer vision tasks. There have been several deep
learning-based palmprint recognition models, which outper-
form other traditional algorithms [5], [11], [12].
Manuscript received March 16, 2021; revised April 19, 2021; accepted However, current palmprint recognition methods based on
April 24, 2021. Date of publication April 30, 2021; date of current version deep learning indeed have some application disadvantages.
May 14, 2021. This work was supported in part by the National Natural One of the major problems is the requirements of large
Science Foundation of China under Grant 61105021, in part by the Nat-
ural Science Foundation of Zhejiang Province under Grant LGF19F030002, amounts of training samples and labels [13], [14]. Generally,
in part by the Natural Science Foundation of Shaanxi Province under collecting enough data is laborious, and labeling them accu-
Grant 2020JM-073, in part by the Fundamental Research Funds for the Central rately is also a heavier work. In addition, although palmprint
Universities under Grant xzy022020051, and in part by the China Scholarship
Council. The Associate Editor coordinating the review process was Hongrui recognition may be more private than face recognition in
Wang. (Corresponding author: Dexing Zhong.) practical application, sometimes users are still reluctant to
Huikai Shao is with the School of Automation Science and provide many images for training and registration due to
Engineering, Xi’an Jiaotong University, Xi’an 710049, China (e-mail:
[email protected]). privacy considerations. Furthermore, even if we have enough
Dexing Zhong is with the School of Automation Science and Engineering, data, training on large amounts of data would be computa-
Xi’an Jiaotong University, Xi’an 710049, China, also with the Pazhou Lab- tionally expensive with poor generalization ability. Therefore,
oratory, Guangzhou 510335, China, and also with the State Key Laboratory
for Novel Software Technology, Nanjing University, Nanjing 210093, China constructing an accurate and effective palmprint recognition
(e-mail: [email protected]). system, which requires less training data, is a significant task.
Xuefeng Du is with the Department of Computer Science, University of The problem reflected above is the so-called small-
Wisconsin–Madison, Madison, WI 53201 USA (e-mail: [email protected]).
Shaoyi Du is with the Institute of Artificial Intelligence and Robotics, Xi’an sample-size (SSS) problem [15], which is also called few-shot
Jiaotong University, Xi’an 710049, China (e-mail: [email protected]). recognition. It is formulated as follows: given a dataset S
Raymond N. J. Veldhuis is with the Faculty of Electrical Engineering, Math- consisting of palmprint images, and for each category, we have
ematics and Computer Science, University of Twente, 7500 AE Enschede, The
Netherlands (e-mail: [email protected]). k samples labeled (k is a small number). The task is how can
Digital Object Identifier 10.1109/TIM.2021.3076850 we recognize the rest images in each category using a little
1557-9662 © 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

zed licensed use limited to: MINISTERE DE L'ENSEIGNEMENT SUPERIEUR ET DE LA RECHERCHE SCIENTIFIQUE. Downloaded on April 24,2023 at 16:15:31 UTC from IEEE Xplore. Restriction
 5009812 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 70, 2021

labeled training data. It is similar to the practical palmprint can outperform others to obtain the state-of-the-art palm-
recognition scenarios, where a few images are registered in print recognition.
the database and the query images need to be matched with Compared with our previous work in [20], we have made
the registration images to determine the tester’s identity. One some significant improvements. First, in addition to the pre-
of the effective solutions is meta-learning [16], which aims vious similarity loss, two other losses are constructed to
to train the deep neural networks (DNNs) for generalizing on constrain the distance of image pairs in the feature space
different tasks. Inspired by it, in this article, we propose a directly, i.e. contrastive loss and BD. Though obtaining
few-shot palmprint recognition method called meta-Siamese the category relations between two images matched through
network (SN) (MSN). neural networks can reduce the impact of manual interven-
In order to help DNN to generalize to new palmprint tion, the distance constraints on them will be beneficial and
images, our MSN follows the structure of SN [17], which improve the performance, which is shown in the results. Here,
firstly extracts feature from image pairs using weight-shared distance-based losses can make positive matching features
convolutional neural network (CNN). Then, the feature vectors closer while negative matching features farther in the feature
extracted are concatenated together and input to follow-up space. Second, eight new unconstrained palmprint databases
decision network to obtain their similarity. Finally, the mean and four benchmark palmprint databases are introduced in
square error (MSE)-based similarity losses are adopted and the experiments to verify the effectiveness of our modified
backpropagated so that it can verify whether palmprint image algorithms. Third, more adequate analyses and comparisons
pairs are from the same individual. are conducted with the state-of-the-art algorithms to demon-
However, different from SN, meta-learning is introduced strate the superiority of our algorithms, especially the recent
to improve the generalization ability and the model is trained few-shot palmprint recognition methods.
through episode-based iteration. Specifically, images of N The remainder of this article is structured as follows.
categories are firstly randomly sampled from train set, and for Section II reviews some related works. Our methods are
each category, k images are selected, which are matched with described in Section III in detail. Section IV presents our
the images in query set (denoted as a N-way, k-shot episode experiments and results on several databases. Analysis of
task). In the testing stage, a similar strategy is adopted to results is in Section V. Section VI gives a conclusion for this
select support set and testing set from test set, where the article.
former is labeled and the latter is unlabeled, and N-way,
k-shot tasks are also randomly sampled. The model is trained II. R ELATED W ORK
on a large number of N-way, k-shot tasks in the training
set, and finally it can adapt the new N-way, k-shot tasks A. Palmprint Recognition
to obtain the similarity scores of different palmprint image Traditional palmprint recognition algorithms mainly
pairs. In order to help optimize the model and improve the extract its rich main line, texture, and wrinkle features.
performance, two distance-based losses, contrastive loss [18] One kind of the commonly used methods is based on
or binomial deviance (BD) [19], are introduced to constrain orientation code. They convolve palmprint images with a
the distance between features in the feature space. Specially, list of Gabor filters with several orientations and convert
to increase the flexibility, the convolutional blocks are them into codes as features, such as competitive code [21],
incorporated in the decision network instead of pure stacked binary orientation co-occurrence vector (BOCV) [22],
fully connected (FC) layers so that the entire neural networks extend BOCV (E-BOCV) [23], double-orientation code
can adapt quickly. Experiments on several popular benchmark (DOC) [24], discriminative and robust competitive code
palmprint datasets reveal the outperforming accuracy and (DRCC) [25], and so on. Using the multiplication and
generalization ability of our model. The details can be found addition schemes, Fei et al. [26] fused the apparent and
in Second III. The overview of MSN is shown in Fig. 1. latent direction features of palmprint and proposed a unique
The contributions can be summarized as follows. double-layer direction extraction method, called apparent
1) MSN is proposed for efficient few-shot palmprint recog- and latent direction code (ALDC). Luo et al. [27] proposed
nition. Its core is to directly imitate the identification task local line directional patterns (LLDP) which operated
of the test in the training phase to improve the accuracy. in local line-geometry space for palmprint recognition.
After the episode-based training on the tasks in training Zhang et al. [6] established a contactless palmprint database
set, the model can be applied to the test set for new and proposed CR_CompCode for palmprint identification
few-shot palmprint recognition tasks. with low computational complexity. Fei et al. [26] extracted
2) MSE-based similarity loss is applied to measure the sim- six discriminant direction binary codes (DDBCs) for each
ilarity scores of palmprint image pairs to determine their pixel of palmprint image and concatenated them as the global
categories. Distance-based losses are further constructed feature vector, called discriminant direction binary palmprint
to assist in training the model and improve the accuracy. descriptor (DDBPD). Toward more accurate direction
3) Adequate experiments are conducted on several con- representations, Jia et al. [28] extracted the direction features
strained and unconstrained benchmark palmprint data- of palmprint on more levels such as multiscale, multidirection
bases. From the results, MSN can obtain promising level, and multiregion. Zhang et al. [6] proposed a unique local
performance and the best accuracy can be up to 100%. descriptor to extract both direction descriptors and thickness
Furthermore, compared with the previous models, MSN features, called local microstructure tetra pattern (LMTrP).

zed licensed use limited to: MINISTERE DE L'ENSEIGNEMENT SUPERIEUR ET DE LA RECHERCHE SCIENTIFIQUE. Downloaded on April 24,2023 at 16:15:31 UTC from IEEE Xplore. Restriction
 SHAO et al.: FEW-SHOT LEARNING FOR PALMPRINT RECOGNITION 5009812

Fig. 1. Overview of our MSN (5-way, 1-shot). Sample and query images are selected from training set randomly to imitate the N -way, k-shot recognition tasks
in test set, where support and testing images are selected randomly during testing. Sample and query images are input into weight-shared feature extractors to
get feature vectors. Then, they are concatenated and input into decision network to obtain similarity score. Similarity loss and distance loss are constructed
to optimize the model. After training, the network can be used for new N -way, k-shot tasks to distinguish the identity of other images in the test set.

However, the methods abovementioned need to design and B. Few-Shot Learning

extract feature manually, which are easily affected by sub-
jective factors, especially for unconstrained palmprint images. CNN uses spatial filters to compute the weighted sum in
Recently, CNN exhibits powerful feature extraction capabil- image patches, which may cause huge computation com-
ity, and deep learning-based palmprint recognition methods plexity and weak generalization ability. As an application
obtain the state-of-the-art performance. Genovese et al. [11] of meta-learning, few-shot learning is introduced to solve
proposed a novel CNN model called PalmNet using princi- the problem of overfitting when CNN is faced with sparse
pal component analysis (PCA) and Gabor responses. Shao data, which samples auxiliary tasks to help DNN general-
and Zhong [29] adopted deep hashing network (DHN) for ize [33]. There exist three types of few-shot learning methods,
palmprint recognition and fused it with dorsal hand vein to namely recurrent neural network (RNN)-based, initialization-
perform multibiometrics, which transferred palmprint images based, and metric-based models. RNN-based methods take the
into binary codes. Meraoumia et al. [30] proposed a novel advantage of memory and train RNN models to remember
deep learning architecture to extract texture information for the seen tasks [34]. When facing tasks with a different
palmprint identification, called PCANet. Shao and Zhong [29] distribution, it compares them with its memory [35]. The
proposed graph neural network (GNN) for few-shot palm- second models aim to help DNN learn to update its parameters
print recognition, where the features extracted by CNN were within a few gradient steps; therefore, the sampled tasks
processed into nodes and the edges were used to represent in training resemble the task setting in testing [33]. The
similarities between images. Matkowski et al. [5] established metric-based models learn the image embedding functions
a new palmprint database collected from uncontrolled and so that the subspace features after embedding are easier to
uncooperative environment and proposed an end-to-end deep classify [36].
learning algorithm for ROI extraction and feature matching. Specifically, Snell et al. [37] proposed prototypical
Zhao and Zhang [31] proposed deep discriminative represen- networks for few-shot learning and computed distances
tation (DDR) to extract high-level discriminative features for to prototype features of every category to perform the
palmprint recognition. Some researchers also proposed algo- classification. Then, Ren et al. [38] modified it and proposed
rithms for cross-dataset recognition, such as PalmGAN [32] a novel few-shot learning algorithm when unlabeled examples
and transfer convolutional autoencoder [29]. were available for producing prototypes. The model was
These deep learning-based palmprint recognition methods trained in an end-to-end way on episodes and learned to
can obtain promising accuracy. However, they require many leverage the unlabeled examples successfully. Sung et al. [33]
labeled palmprint images, which consumes a lot of manpower proposed relation network (RN) to learn a deep distance
and material resources. Different from them, in this article, we metric to compare images and classify new classes by
focus on few-shot palmprint recognition and propose a novel computing their relation scores. Finn et al. [39] proposed a
MSN method, which can achieve good performance based on model-agnostic meta-learning algorithm, and the parameters
a few labeled data. could be explicitly trained using a few data for a new task.

zed licensed use limited to: MINISTERE DE L'ENSEIGNEMENT SUPERIEUR ET DE LA RECHERCHE SCIENTIFIQUE. Downloaded on April 24,2023 at 16:15:31 UTC from IEEE Xplore. Restriction
 5009812 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 70, 2021

Kang et al. [40] proposed a few-shot object detector using

a meta-feature extractor and a reweighting learner within
a one-stage detection structure. Recently, researchers also
proposed GNN and knowledge graph-based algorithms for
few-shot learning and achieved promising results, such as
graph few-shot learning (GFL) [41], GNN [42], hybrid
knowledge routed modules (HKRM) [43], and so on.
Liu et al. [44] proposed a novel transductive propagation
network (TPN) for few-shot classification task. They proposed
to learn a graph construction module to exploit the manifold
structure in the data and classify the entire test set at once. Fig. 2. Pipeline of 5-way, 1-shot palmprint recognition. Five images of five
categories are randomly selected to be matched with the query images, and
Wang et al. [45] proposed a novel dubbed instance credibility then their similarity scores are obtained.
inference (ICI) method to perform the distribution support
of unlabeled instances for few-shot learning. Ravi and Beat-
son [46] proposed a novel meta-learning method to amortize set, {(x i , yi ), i = N × k + 1, . . . , M}. Finally, N-way k-shot
hierarchical variational inference across tasks so that a few tasks are formed to train the model in an episode manner.
steps of Bayes by backprop can produce useful task-specific In testing, the samples are randomly selected from test set to
approximate posteriors. Li et al. [47] proposed TargetNet form the support and testing sets, which is similar to training
and MetaNet to learn transferable knowledge across different set but the support set has labels and the testing set has no
tasks, which can construct parameters for similar unseen tasks. labels. The losses obtained on the sample and query sets are
The improvement of our work compared with [17] and [48] backpropagated to adapt the model to new tasks. Therefore,
is that we add meta-episode training to improve the general- after the episode-based training, the model can be applied to
ization ability. Besides, we add CNN blocks in the decision the test set for new few-shot palmprint recognition tasks.
network which enables more flexible training. Compared with
RN [33], we introduced two distance-based losses to help opti-
mize the feature distributions, which can achieve promising B. One-Shot Palmprint Recognition
improvement on the performance. Compared with [35], our For one-shot learning, there is only one labeled image for
model does not require designing complex memory encoding each category in the sample set and support set. A schematic
and decoding structure. of 5-way, 1-shot palmprint recognition is shown in Fig. 2.
During training, firstly, N images from N categories (one
III. M ETHODS image in each category) in training set are randomly sampled
to form the sample set {(x 00 , y00 ), (x 01 , y01 ), . . . , (x 0N −1 , y0N −1 )},
A. Task Description and the remaining images of the same N classes are denoted
j
Few-shot palmprint recognition can be formulated as train- as the query set {(x ii , yi ), j = 0, . . . , N − 1; i = 1, . . . , q}.
ing a classifier to recognize the remaining images given a Here, we suppose each class consists of q images in total.
few labeled images for each category in the test set. Due to During every episode-based iteration, one image, x 0s , in the
j
lack of enough labeled data, if the classifier is directly trained sample set and one image, x i , in the query set are input
using traditional optimization algorithms (e.g. softmax loss), to two CNN feature extractor modules f (·), respectively,
it will suffer from overfitting and may not obtain satisfactory which have the same weights, to obtain two feature vectors,
j
performance. So the meta-learning is proposed [16]. Suppose f (x 0s ) and f (x i ). Then, they are concatenated together in
there is an image dataset, {(x 0 , y0 ), . . . , (x i , yi ), . . . , (x n , yn )}, the depth dimension to obtain the concatenated feature map,
j
and yi is the label of x i . The dataset is split into training set C( f (x 0s ), f (x i )). Finally, by down-sampling and dimension
and test set. The test set is further divided into support set and transformation in the decision network D(·), a similarity score
testing set, and we aim to match the images of testing set to the with respect to the current image pair is calculated. When
support set accurately. It is similar to palmprint identification the score is higher, it means the two images are more likely
scenario, where the test images are matched to the registration to belong to the same individual. Therefore, for one episode
samples in the database to determine their identity. During training iteration, the similarities between different image pairs
training, the training set is separated to sample set and query are compared for N × N × (q − 1) times.
set to simulate the few-shot recognition setting in the test set. For parameter optimization, an MSE-based similarity loss
Note that the label space of the sample/query sets is different is adopted. A one-hot vector, Y , is constructed to denote the
from that of support/testing sets. In addition, the images and ground-truth value, which is a N × N × (q − 1) dimensional
labels are not used at all during training. vector. The t-th value of Y is formulated as follows:
Similar to [49], in each episode-based training itera- 
1, if t-thimage pair is positive
tion, M images of N categories are randomly selected Yt = (1)
from training set, and N × k < M images are fur- 0, otherwise.
ther randomly selected to denote the small sample set as The predicted similarity score is set as
{(x i , yi ), i = 1, . . . , N × k} for training, where each class has      j 
k samples. Then, the remaining images are denoted as query Ŷt = D C f x 0s , f x i . (2)

zed licensed use limited to: MINISTERE DE L'ENSEIGNEMENT SUPERIEUR ET DE LA RECHERCHE SCIENTIFIQUE. Downloaded on April 24,2023 at 16:15:31 UTC from IEEE Xplore. Restriction
 SHAO et al.: FEW-SHOT LEARNING FOR PALMPRINT RECOGNITION 5009812

Then, the similarity loss is adopted

N ×N ×(q−1)

Ls = (Yt − Ŷt )2 . (3)
t=1
Though the loss above can be optimized to constrain
the similarity score of positive matching approaching 1 and
the score of negative matching approaching 0, the score is
obtained through decision network, which needs to be care-
fully designed and takes some time to train. Therefore, inspired
by metric learning, another distance-based loss is constructed
for auxiliary optimization, which can be contrastive loss [18]
or BD [19].
j
For images x 0s and x i , the contrastive loss is
 j     j 
L c x 0s , x i = S0i d f x 0s , f x i Fig. 3. Pipeline of few-shot recognition (5-way, 3-shot). For every category,
     j   three images are selected as sample images to get the features, which are
+ (1− S0i ) max m − d f x 0s , f x i , 0 (4) averaged to a feature map. Then, the averaged feature maps are matched with
j query features to obtain the similarity scores.
where d( f (x 0s ), f (x i )) is the Euclidean distance between
j
f (x 0s ) and f (x i ), and m is a distance margin. S0i is the image
pair label and set to 1 for positive image pairs and 0 for
negative image pairs.
For BD, it is
 j
L BD x 0s , x i = log(1 + e−(2c0i −1)β1 c0i −β2 )α ) (5)
 s T  j 
f x f x
c0i =   0  i j  (6)
 f x s  f x 
0 i
where c0i is their cosine similarity score. β1 and β2 are scaling
and translation parameters and set to 2 and 0.5, like [50]; α
is a balance weight and set to 1 for positive image pairs and
25 for negative image pairs.
Through these two losses, the features of positive matching
can be as close as possible and the features of negative
matching can be far away.
Therefore, the overall loss is Fig. 4. Structure of two sub-modules, namely CNN feature extractor and
decision network. The configuration of each layer is shown in the box. For
L = L s + wL d (7) instance, the “64 × 3 × 3.1-0” means there are 64 convolutional filters with
3 × 3 kernel size, 1 stride, and 0 padding. “FC layer (128)” means we have
where w is a parameter to balance the weights of two losses 128 hidden units in the FC layer. Activation functions and batch normalization
and L d can be L c or L BD . are also demonstrated in each box.
At the testing stage, the support set and testing set are
constructed in a similar way to the sample set and query set.
map is obtained. Afterward, the averaged feature vector is
For each of the N × (q − 1) testing images, it is compared
concatenated with the features from the query images and
with every image in the support set, and the image in the
input to decision network to get similarity score. Similarly,
support set which generates the highest similarity score is
the distance between the averaged feature vector and query
selected. Then, the label of support image selected is compared
feature is also constrained through distance-based losses. After
with the label of testing image to evaluate whether our model
that, the following matching and comparison procedures are
generates the right recognition result, and the accuracy can be
the same as the one-shot palmprint recognition. The diagram
calculated.
of few-shot palmprint recognition scenario is shown in Fig. 3.
In this article, CNN feature extractor is adopted to obtain
C. k-Shot Palmprint Recognition
convolutional features and decision network is adopted to get
For k-shot palmprint recognition, similar to one-shot recog- similarity score. Their parameters are shown in Fig. 4.
nition, k images are sampled instead of one image for each
category in sample set and the remaining images are regard IV. E XPERIMENTS AND R ESULTS
as query images. The difference is that we input k sample
images along with query images to MSN at each episode A. Databases
training. In order to help the neural networks to see more PolyU multispectral palmprint database consists of four
information of a certain class, all k feature maps of sam- spectral bands, i.e. blue, red, green, and near-infrared
ple images are summed up and then their averaged feature (NIR) [51]. Under each spectral band, there are 6000 images

zed licensed use limited to: MINISTERE DE L'ENSEIGNEMENT SUPERIEUR ET DE LA RECHERCHE SCIENTIFIQUE. Downloaded on April 24,2023 at 16:15:31 UTC from IEEE Xplore. Restriction
 5009812 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 70, 2021

TABLE I
S OME D ETAILS OF D IFFERENT PALMPRINT D ATABASES

Fig. 5. Some typical ROI samples of PolyU multispectral palmprint database.

(a) Blue, (b) green, (c) red, and (d) NIR.

which consists of 1620 image from 81 hands. CASIA

Fig. 6. Some typical samples of Tongji palmprint database. (a) and palmprint database contains 5502 palm images collected
(b) Original images. (c) ROI.
from 312 individuals. There are 2601 images collected
from 230 individuals in IITD palmprint database. All of
collected from 500 different palms of 250 individuals, includ- them are cropped to form the ROIs using official methods.
ing 195 males and 55 females. For each palm, 12 images After preprocessing, five images are selected for each palm.
were collected in two sessions, and six images in each session. The details of these palmprint databases are summarized
The palmprint images are oriented and cropped to form ROIs in Table I.
with the size of 128 × 128 pixels. According to different
spectral bands, this database can be used as four sub-databases, B. Implementation Details
recorded as blue, red, green, and NIR. Fig. 5 shows some
typical samples. In this article, the image dataset is randomly split into
the training set and test set with the ratio of 1:1. The model
Tongji contactless palmprint database contains
12 000 images captured from 300 individuals [6]. For is evaluated in 15 different constrained and unconstrained
palmprint sub-datasets on different task settings, which are
each category, there are 20 palm images. During acquisition,
5-way/1-shot, 5-way/3-shot, 5-way/5-shot, 15-way/1-shot,
the volunteers can move their hands up and down freely
within the enclosed space. A ring white LED light source 15-way/3-shot, and 15-way/5-shot recognition tasks. Images
are resized to 128 × 128 and input to the networks. The entire
was used and the background was still pure black, which
experiments are implemented using PyTorch framework on
reduces the difficulty of recognition. Some examples are
shown in Fig. 6, and ROIs with the size of 128 × 128 pixels NVIDIA GPU GTX 1080 and i7-3.30 GHz processors. The
base learning rate is set to 0.001, and the Adam Optimizer
are extracted based on [6].
Xi’an Jiaotong University Unconstrained Palmprint (XJTU- and Stochastic Gradient Descent (SGD) are adopted.
UP) database was established using five mobile phones, i.e.,
Samsung Galaxy Note5, iPhone 6S, LG G4, HUAWEI Mate8, C. Results
and MI8 [52], [53]. The volunteers can freely move their hands 1) Performance on PolyU Multispectral Palmprint Data-
during acquisition and the backgrounds are complex natural base: For each sub-database, 3000 images from 250 categories
scenes. Two kinds of illuminations were adopted, indoor nat- are used as training set and the remaining images are used as
ural illumination and the flash lighting of mobile phone. There test set. For 5-way, 1-shot recognition scenario, the size of
are ten sub-databases in XJTU-UP database, named HUAWEI sample set is 5 and the size of query set is 55. For 5-way,
Mate8 under Natural illumination (HN), iPhone 6s under 3-shot recognition scenario, the size of sample set is 15 and
Natural illumination (IN), LG G4 under Natural illumination the size of the query set is 45. For 5-way, 5-shot recognition
(LN), MI8 under Natural illumination (MN), Samsung Galaxy scenario, the size of sample set is 25 and the size of query set is
Note5 under Natural illumination (SN), HUAWEI Mate8 under 35. Similarly, for 15-way, 1-shot recognition scenario, the size
Flash illumination (HF), iPhone 6s under Flash illumination of sample set is 15 and the size of query set is 165. The results
(IF), LG G4 under Flash illumination (LF), MI8 under Flash on different experiment settings are shown in Table II. The top
illumination (MF), and Samsung Galaxy Note5 under Flash 1 accuracy is marked in bold (same below). “MSE” means
illumination (SF). Each sub-database contains RGB hand the overall loss L = L s , “Con” means L = L s + wL c , and
images from 100 individuals and each category has ten images. “BD” means L = L s + wL BD , and the description below is
The ROIs are extracted based on the key points and methods similar, unless otherwise specified. From the results, for Blue,
provided by [52]. Fig. 7 shows some exemplary samples of the optimal accuracy is 100% on 5-way, 5-shot recognition
XJTU-UP database. using similarity loss and contrastive loss. For Green, the opti-
Furthermore, three other palmprint databases are also mal accuracy is 100% both on 5-way, 3-shot and 5-way, 5-
adopted to evaluate the model, NUIG_Palm1 [54], shot recognition using similarity loss and BD loss. For NIR,
CASIA [55], and IITD [56]. NUIG_Palm1 database is the optimal accuracy is 100% on 5-way, 5-shot recognition
also collected by mobile phones under different conditions, using similarity loss and BD loss. For Red, the accuracies on

zed licensed use limited to: MINISTERE DE L'ENSEIGNEMENT SUPERIEUR ET DE LA RECHERCHE SCIENTIFIQUE. Downloaded on April 24,2023 at 16:15:31 UTC from IEEE Xplore. Restriction
 SHAO et al.: FEW-SHOT LEARNING FOR PALMPRINT RECOGNITION 5009812

TABLE II TABLE IV
F EW-S HOT R ECOGNITION A CCURACIES (%) ON M ULTISPECTRAL F EW-S HOT R ECOGNITION A CCURACIES (%) ON XJTU-UP D ATABASE
D ATABASE

TABLE III
F EW-S HOT R ECOGNITION A CCURACIES (%) ON T ONGJI PALMPRINT
D ATABASE

5-way, 3-shot using similarity loss and 15-way, 5-shot recogni-

tion using similarity loss and BD loss are 100%. In summary,
almost all results of different settings exceed 99%.
2) Performance on Tongji Contactless Palmprint Data-
base: For Tongji palmprint database, 6000 images from
300 categories are selected as training set and the remaining
6000 images are as test set. Each category has 20 images,
so for 5-way, 1-shot recognition scenario, the size of sample 3-shot recognition scenario, the size of sample set is 15 and the
set is 5 and the size of query set is 95. For 5-way, 3-shot size of query set is 35. For 5-way, 5-shot recognition scenario,
recognition scenario, the size of sample set is 15 and the size the size of sample set is 25 and the size of query set is also 25.
of query set is 85. For 5-way, 5-shot recognition scenario, Similarly, for 15-way, 1-shot recognition scenario, the size of
the size of sample set is 25 and the size of query set is 75. sample set is 15 and the size of query set is 135. For 15-way,
For 15-way, 1-shot recognition scenario, the size of sample set 3-shot recognition scenario, the size of sample set is 45 and
is 15 while the size of query set is 285. For 15-way, 3-shot the size of query set is 105. For 15-way, 5-shot recognition
recognition scenario, the size of sample set is 45 and the size scenario, the size of sample set is 75 and the size of query set
of query set is 255. For 15-way, 5-shot recognition scenario, is 75. The results are listed in Table IV.
the size of sample set is 75 and the size of query set is 225. The For IF, the optimal accuracy is 99.954% on 5-way, 5-shot
results are presented in Table III and the optimal accuracy is recognition using similarity loss and contrastive loss. For IN,
99.991% on 15-way, 5-shot recognition using similarity loss the optimal accuracy is 99.909% on 5-way, 5-shot recognition
and contrastive loss. Though the database is collected in a using similarity loss and BD loss. For HF, the optimal
contactless manner, its performance is also relatively good. accuracy is 99.996% on 15-way, 5-shot recognition using
3) Performance on XJTU-UP Database: XJTU-UP data- similarity loss and BD loss. For HN, the optimal accuracy is
base consists of ten sub-databases. For each sub-database, 99.915% on 5-way, 5-shot recognition using similarity loss
the first 100 categories are used as training set and the and contrastive loss. For LF, the optimal accuracy is 99.975%
remaining images are used for testing. Each category contains on 5-way, 5-shot recognition using similarity loss and BD loss.
ten images, and for 5-way, 1-shot recognition scenario, the size For LN, the optimal accuracy is 99.907% on 5-way, 5-shot
of sample set is 5, while the size of query set is 45. For 5-way, recognition using similarity loss. For MF, the optimal accuracy

zed licensed use limited to: MINISTERE DE L'ENSEIGNEMENT SUPERIEUR ET DE LA RECHERCHE SCIENTIFIQUE. Downloaded on April 24,2023 at 16:15:31 UTC from IEEE Xplore. Restriction
 5009812 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 70, 2021

TABLE V
F EW-S HOT R ECOGNITION A CCURACIES (%) ON HF D ATABASE W ITH
D IFFERENT H YPERPARAMETERS

Fig. 8. Results of different hyperparameters. (a) and (b) “Con.” (c) and
(d) “BD.” In each subgraph, the horizontal axis represents w, and the vertical
axis represents the recognition accuracy.

TABLE VI
F EW-S HOT R ECOGNITION A CCURACIES (%) ON HF D ATABASE W ITH
Fig. 7. Some typical samples of XJTU-UP database. (a) and (b) Original
images in HF and IN. (c) and (d) ROIs in HF and HN. D IFFERENT L OSSES

is 99.759% on 5-way, 5-shot recognition using similarity loss

and contrastive loss. For MN, the optimal accuracy is 99.905%
on 5-way, 5-shot recognition using similarity loss and BD loss.
For SF, the optimal accuracy is 99.955% on 5-way, 5-shot
recognition using similarity loss and contrastive loss. For SN,
the optimal accuracy is 99.920% on 5-way, 3-shot recognition
using similarity loss. The results of datasets collected in flash
light are better than that of databases collected in natural
illumination. The flash light can reduce the influence of
external light so that the texture and wrinkles are clearer.
TABLE VII
F EW-S HOT R ECOGNITION A CCURACIES (%) ON HF D ATABASE U SING
D. Ablation Study D IFFERENT D ECISION N ETWORKS
1) Effect of Hyperparameter: Here, we conducted sev-
eral experiments on HF database to show the effect of w,
which is used to balance the weights of similarity loss and
distance-based loss. We set different values for w and the
results are in Table V and Fig. 8. From the results, in most
cases, as w increases, the accuracy increases first and then
decreases. Though the distance-based losses can improve the
performance, their weights cannot be too heavy, which shows
the effectiveness of the flexible decision network. the performance is worst, while combined losses can obtain
2) Roles of Different Losses: Different losses are adopted better accuracy. In addition, the accuracy of “MSE” is higher
to optimize the model. In this part, we conducted several than that of distance-based losses on some datasets. It means
experiments to verify the roles that different sub-losses play. the distance-based losses may bring a negative impact, but
HF is also selected, and different losses and their combinations such scenarios are rare.
are evaluated. Table VI shows the results, where “Single 3) Necessity of Convolutional Blocks in Decision Network:
Con” means only the contrastive loss is adopted, “Single In this article, convolutional blocks are constructed in the deci-
BD” means only BD is used, “MSE + 0.5 × Con” means sion network to increase the flexibility, instead of pure-stacked
the similarity loss is combined with contrastive loss, and FC layers. Here, several experiments are conducted on HF
“MSE + 0.5 × BD” means the similarity loss is combined database to show the necessity of convolutional blocks. The
with BD loss. “MSE + 0.5 × Con + 0.5 × BD” means the results are shown in Table VII. The decision network with
similarity loss is combined with both contrastive loss and or without convolutional blocks is adopted, and other experi-
BD loss as a joint loss. The results are shown in Table VI. ment settings are consistent. Form the results, convolutional
It can be observed that the accuracy of joint loss is higher blocks can extract more potential features for much better
than that of single loss while lower than “MSE + 0.5 × Con” performance, though only the 128-dimensional features are
or “MSE + 0.5 × BD.” If there is only the BD adopted, extracted.

zed licensed use limited to: MINISTERE DE L'ENSEIGNEMENT SUPERIEUR ET DE LA RECHERCHE SCIENTIFIQUE. Downloaded on April 24,2023 at 16:15:31 UTC from IEEE Xplore. Restriction
 SHAO et al.: FEW-SHOT LEARNING FOR PALMPRINT RECOGNITION 5009812

V. E VALUATION AND A NALYSES 5) DHN [57] converts palmprint images into binary codes,
which can improve the efficiency of authentication and
A. Comparisons Between Different Settings
obtain the state of the arts on the traditional palmprint
1) Comparisons Between Different “Ways” and “Shots”: recognition scenario.
In the experiments, 5-way/1-shot, 5-way/3-shot, 5-way/5-shot, 6) GNN [29] uses nodes to represent image features and
15-way/1-shot, 15-way/3-shot, and 15-way/5-shot recognition edge to represent their positive or negative relation.
tasks are performed. From the results, the accuracies of 5-way 7) DRCC [25] adopts a more accurate dominant orientation
are better than that of 15-way. In this article, meta-learning representation of palmprint by weighting the orientation
strategies are adopted. In each episode-based training iteration, information of a neighbor area.
it seems that the network tries to classify the images in the 8) ALDC [58] extracts the apparent and latent direction fea-
sample/query sets. In 15-way few-shot recognition, the knowl- tures and pools them as the histogram feature descriptor.
edge acquired by sampling tasks when training may not be 9) LDDBP [59] adopts a novel exponential and Gaussian
specific enough for handling classification problem among fusion model (EGM) to present the discriminative power
larger amounts of classes. The categories in 5-way setting of different directions of palmprint.
are less, so the classification task is easier and the accuracy 10) DDBPD [26] concatenates several binary feature DDBC
is higher. Furthermore, if there are more labeled images in codes as a global feature vector to perform recognition.
sample set (more “shots”), most of the results will be also 11) PCANet [30], [60] applies cascaded PCA, binary hash-
better, because more knowledge of a certain class is obtained ing, and block-wise histograms to extract features.
by MSN. It can also be observed that the accuracy of 3-shot 12) PalmNet [11] combines Gabor responses and CNN, and
is higher than that of 5-shot in some datasets, such as Red is trained by an unsupervised procedure.
in Table II. It may be because there is more variation between 13) TPN [44] learns a graph construction module to prop-
images, and more “shots” increase the difficulty of learning. agate labels from labeled instances to unlabeled test
2) Comparisons Between Different Databases: In this instances.
article, three benchmarks are adopted. PolyU multispectral 14) LGM-Net [47] learns transferable knowledge across
database and Tongji contactless database are constrained data- different tasks and produces network parameters for
bases, which are collected in closed space with additional similar unseen tasks through TargetNet and MetaNet.
illuminations. So their qualities are better and easier to 15) ABLM [46] amortizes hierarchical variational inference
identify. XJTU-UP database consists of images collected by across tasks and learns a prior distribution over neural
mobile phones in an unconstrained manner, so they contain network weights.
more noise. From the results, the performances of constrained 16) LS loss [61] samples an equal number of negative pairs
images outperform the unconstrained images, but the latter as the positive pairs to take full advantage of training
are also good. However, the unconstrained acquisition is more batches.
suitable for mobile terminal application scenarios. 17) MS loss [62] adopts two iterative steps with sampling
and weighting to improve the performance.
B. Comparisons With Other Models
Note that all of modules are implemented using similar
For comparison, we present the results of some base- hyperparameters with a slight difference in each model to
line methods in the 5-way, 1-shot recognition, namely reach the best performance, respectively. LS loss and MS loss
SN [17], model-agnostic meta-learning (MAML) [39], Pro- are adopted to train deep metric model to extract discriminative
totypical Nets (P-Net) [37], DHN [57], Matching Net (M- features, and Resnet 18 is used as the backbone [63]. The
Net) [49], GNN [29], DRCC [25], ALDC [58], local discrim- experiment settings are kept as consistent, such as the split
inant direction binary pattern (LDDBP) [59], DDBPD [26], of the training data and test data and evaluation method.
PCANet [30], [60], PalmNet [11], TPN [44], LGM-Net [47], From the tables, our model can achieve competitive results
ABLM [46], lifted structure (LS) loss [61], and multisimilar- compared with several popular low-shot recognition methods,
ity (MS) loss [62]. The results are shown in Tables VIII and namely M-Net, P-Net, MAML, GNN, TPN, LGM-Net, and
IX, and the top 1 accuracy is highlighted with bold. ABLM. Compared with the state-of-the-art palmprint recogni-
1) SN [17] uses two same networks to extract features and a tion models using traditional training strategies without special
decision network to get the similarity scores of matched design for low-shot recognition, our model performs better
images. in all datasets. ALDC, LDDBP, and DDBPD are handcrafted
2) MAML [39] is based on meta-learning and aims to palmprint recognition methods. Though they can obtain rela-
explicitly train a network on a number of learning tasks tively high accuracy on constrained database, but they are not
so that it can adapt to new learning tasks. as good as deep learning-based methods, such as PalmNet.
3) P-Net [37] adopts distance-based loss to learn a metric LS loss and MS loss can obtain satisfactory performance when
space and achieves classification by obtaining the dis- there are enough training data. However, compared with MSN,
tances to prototype representations of every category. due to the lack of labeled training data, their performances are
4) M-Net [49] is based on deep metric learning and aug- also limited. SN and DHN are supervised algorithms, but there
ments neural networks with external memories to adapt are not enough labeled samples here, so their performances
new tasks. have dropped significantly. In the SSS palmprint recognition

zed licensed use limited to: MINISTERE DE L'ENSEIGNEMENT SUPERIEUR ET DE LA RECHERCHE SCIENTIFIQUE. Downloaded on April 24,2023 at 16:15:31 UTC from IEEE Xplore. Restriction
 5009812 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 70, 2021

TABLE VIII
C OMPARATIVE R ESULTS (%) OF F EW-S HOT R ECOGNITION ON D IFFERENT M ODELS

TABLE IX
C OMPARATIVE R ESULTS (%) OF F EW-S HOT R ECOGNITION ON D IFFERENT M ETHODS

scenario (only a few labeled samples can be used for train- VI. C ONCLUSION
ing and registration), which is more common in practical In this article, a novel few-shot model, MSN, is proposed for
applications, these previous methods do not work well, and palmprint recognition only using a few labeled images. On the
this shows the effectiveness of our proposed methods. From basis of classical SN, meta-episode training is introduced
the results, few-shot learning-based methods can generally for better generalization performance. Two weight-shared net-
achieve better performance, especially for constrained data- works are adopted to extract the features, which are measured
bases. However, our MSN combined similarity loss with by a decision network to obtain their similarities. In the
distance loss can obtain higher accuracy on few-shot palmprint similarity metric learning stage, the initial model is modified
recognition. to compare two palmprint images flexibly by introducing CNN

zed licensed use limited to: MINISTERE DE L'ENSEIGNEMENT SUPERIEUR ET DE LA RECHERCHE SCIENTIFIQUE. Downloaded on April 24,2023 at 16:15:31 UTC from IEEE Xplore. Restriction
 SHAO et al.: FEW-SHOT LEARNING FOR PALMPRINT RECOGNITION 5009812

blocks. Inspired by meta-learning, the training set is split into [17] D. Zhong, Y. Yang, and X. Du, “Palmprint recognition using siamese
sample/query sets and the test set is split into support/testing network,” in Proc. Chin. Conf. Biometric Recognit., Urumqi, China,
2018, pp. 48–55.
sets. Furthermore, two distance-based losses are adopted to [18] S. Chopra, R. Hadsell, and Y. LeCun, “Learning a similarity metric
assist the optimization, which makes the positive matchings discriminatively, with application to face verification,” in Proc. IEEE
closer and negative matchings farther in the feature space. Comput. Soc. Conf. Comput. Vis. Pattern Recognit. (CVPR), San Diego,
CA, USA, Jun. 2005, pp. 539–546.
Finally, the model learns the ability to measure the similarity [19] D. Yi, Z. Lei, S. Liao, and S. Z. Li, “Deep metric learning for person
between two palmprint images on the learning tasks during re-identification,” in Proc. 22nd Int. Conf. Pattern Recognit., Stockholm,
training, which can adapt the new recognition tasks in the Sweden, Aug. 2014, pp. 1–11.
[20] X. Du, D. Zhong, and P. Li, “Low-shot palmprint recognition based
test set. Experiments on several benchmarks, including con- on meta-siamese network,” in Proc. IEEE Int. Conf. Multimedia Expo
strained and unconstrained palmprint databases, show that our (ICME), Shanghai, China, Jul. 2019, pp. 79–84.
algorithms can outperform other methods to be the state of the [21] F. Yue, W. Zuo, D. Zhang, and K. Wang, “Orientation selection
using modified FCM for competitive code-based palmprint recognition,”
arts, and the accuracies can be up to 100%. And, our model is Pattern Recognit., vol. 42, no. 11, pp. 2841–2849, Nov. 2009.
very suitable for hand-based practical personal authentication [22] Z. Guo, D. Zhang, L. Zhang, and W. Zuo, “Palmprint verification using
scenarios when the size of the training or registration set is binary orientation co-occurrence vector,” Pattern Recognit. Lett., vol. 30,
no. 13, pp. 1219–1227, Oct. 2009.
small or when only a small part of palmprint images are
[23] L. Zhang, H. Li, and J. Niu, “Fragile bits in palmprint recognition,”
labeled in the acquisition stage. In the future, we will extend IEEE Signal Process. Lett., vol. 19, no. 10, pp. 663–666, Oct. 2012.
our model to zero-shot recognition scenario by introducing the [24] L. Fei, Y. Xu, W. Tang, and D. Zhang, “Double-orientation code
semantic features of palmprint. and nonlinear matching scheme for palmprint recognition,” Pattern
Recognit., vol. 49, pp. 89–101, Jan. 2016.
[25] Y. Xu, L. Fei, J. Wen, and D. Zhang, “Discriminative and robust
competitive code for palmprint recognition,” IEEE Trans. Syst., Man,
R EFERENCES Cybern. Syst., vol. 48, no. 2, pp. 232–241, Feb. 2018.
[1] A. K. Jain, A. Ross, and S. Prabhakar, “An introduction to biometric [26] L. Fei, B. Zhang, Y. Xu, Z. Guo, J. Wen, and W. Jia, “Learning
recognition,” IEEE Trans. Circuits Syst. Video Technol., vol. 14, no. 1, discriminant direction binary palmprint descriptor,” IEEE Trans. Image
pp. 4–20, Jan. 2004. Process., vol. 28, no. 8, pp. 3808–3820, Aug. 2019.
[2] M. Kopaczka, R. Kolk, J. Schock, F. Burkhard, and D. Merhof, “A [27] Y.-T. Luo et al., “Local line directional pattern for palmprint recogni-
thermal infrared face database with facial landmarks and emotion tion,” Pattern Recognit., vol. 50, pp. 26–44, Feb. 2016.
labels,” IEEE Trans. Instrum. Meas., vol. 68, no. 5, pp. 1389–1401, [28] W. Jia et al., “Palmprint recognition based on complete direction repre-
May 2019. sentation,” IEEE Trans. Image Process., vol. 26, no. 9, pp. 4483–4498,
[3] K. Cao and A. K. Jain, “Automated latent fingerprint recognition,” Sep. 2017.
IEEE Trans. Pattern Anal. Mach. Intell., vol. 41, no. 4, pp. 788–800, [29] H. Shao and D. Zhong, “Few-shot palmprint recognition via graph neural
Apr. 2019. networks,” Electron. Lett., vol. 55, no. 16, pp. 890–891, Aug. 2019.
[4] A.-S. Ungureanu, S. Salahuddin, and P. Corcoran, “Toward uncon- [30] A. Meraoumia, F. Kadri, H. Bendjenna, S. Chitroub, and A. Bouridane,
strained palmprint recognition on consumer devices: A literature review,” “Improving biometric identification performance using PCANet deep
IEEE Access, vol. 8, pp. 86130–86148, 2020. learning and multispectral palmprint,” in Biometric Security and Privacy.
[5] W. M. Matkowski, T. Chai, and A. W. K. Kong, “Palmprint recogni- Cham, Switzerland: Springer, 2017, pp. 51–69.
tion in uncontrolled and uncooperative environment,” IEEE Trans. Inf. [31] S. Zhao and B. Zhang, “Deep discriminative representation for generic
Forensics Security, vol. 15, pp. 1601–1615, 2020. palmprint recognition,” Pattern Recognit., vol. 98, pp. 1–11, Feb. 2020.
[6] L. Zhang, L. Li, A. Yang, Y. Shen, and M. Yang, “Towards contactless [32] H. Shao, D. Zhong, and Y. Li, “PalmGAN for cross-domain palmprint
palmprint recognition: A novel device, a new benchmark, and a collab- recognition,” in Proc. IEEE Int. Conf. Multimedia Expo (ICME), Shang-
orative representation based identification approach,” Pattern Recognit., hai, China, Jul. 2019, pp. 1390–1395.
vol. 69, pp. 199–212, Sep. 2017. [33] F. Sung, Y. Yang, L. Zhang, T. Xiang, P. H. S. Torr, and
[7] L. Fei, B. Zhang, S. Teng, Z. Guo, S. Li, and W. Jia, “Joint multiview T. M. Hospedales, “Learning to compare: Relation network for few-shot
feature learning for hand-print recognition,” IEEE Trans. Instrum. Meas., learning,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., Salt
vol. 69, no. 12, pp. 9743–9755, Dec. 2020. Lake City, UT, USA, Jun. 2018, pp. 1199–1208.
[8] N. B. Mahfoudh, Y. B. Jemaa, and F. Bouchhima, “A robust palmprint [34] A. Santoro, S. Bartunov, M. Botvinick, D. Wierstra, and T. Lillicrap,
recognition system based on both principal lines and Gabor wavelets,” “Meta-learning with memory-augmented neural networks,” in Proc. Int.
Int. J. Image, Graph. Signal Process., vol. 5, no. 7, pp. 1–8, 2013. Conf. Mach. Learn., New York, NY, USA, 2016, pp. 1842–1850.
[9] D. Zhong, X. Du, and K. Zhong, “Decade progress of palmprint [35] T. Munkhdalai and H. Yu, “Meta networks,” in Proc. 34th Int. Conf.
recognition: A brief survey,” Neurocomputing, vol. 328, pp. 16–28, Mach. Learn., Sydney, NSW, Australia, 2017, pp. 2554–2563.
Feb. 2019. [36] L. Bertinetto, J. F. Henriques, J. Valmadre, P. H. S. Torr, and A. Vedaldi,
[10] A. Kong, D. Zhang, and M. Kamel, “A survey of palmprint recognition,” “Learning feed-forward one-shot learners,” in Proc. 30th Conf. Neural
Pattern Recognit., vol. 42, no. 7, pp. 1408–1418, Jul. 2009. Inf. Process. Syst. (NIPS), Barcelona, Spain, 2016, pp. 523–531.
[11] A. Genovese, V. Piuri, K. N. Plataniotis, and F. Scotti, “PalmNet: Gabor- [37] J. Snell, K. Swersky, and R. S. Zemel, “Prototypical networks for few-
PCA convolutional networks for touchless palmprint recognition,” IEEE shot learning,” in Proc. Adv. Neural Inf. Process. Syst., Long Beach,
Trans. Inf. Forensics Security, vol. 14, no. 12, pp. 3160–3174, Dec. 2019. CA, USA, 2017, pp. 4080–4090.
[12] H. Shao and D. Zhong, “Towards cross-dataset palmprint recognition via [38] M. Ren et al., “Meta-learning for semi-supervised few-shot classifica-
joint pixel and feature alignment,” IEEE Trans. Image Process., vol. 30, tion,” in Proc. 6th Int. Conf. Learn. Represent. (ICLR), Vancouver, BC,
pp. 3764–3777, 2021, doi: 10.1109/TIP.2021.3065220. Canada, 2018, pp. 1–15.
[13] C. Yan, B. Shao, H. Zhao, R. Ning, Y. Zhang, and F. Xu, “3D room [39] C. Finn, P. Abbeel, and S. Levine, “Model-agnostic meta-learning for
layout estimation from a single RGB image,” IEEE Trans. Multimedia, fast adaptation of deep networks,” in Proc. 34th Int. Conf. Mach. Learn.,
vol. 22, no. 11, pp. 3014–3024, Nov. 2020. Sydney, NSW, Australia, 2017, pp. 1126–1135.
[14] H. Shao and D. Zhong, “One-shot cross-dataset palmprint recogni- [40] B. Kang, Z. Liu, X. Wang, F. Yu, J. Feng, and T. Darrell, “Few-shot
tion via adversarial domain adaptation,” Neurocomputing, vol. 432, object detection via feature reweighting,” in Proc. IEEE/CVF Int. Conf.
pp. 288–299, Apr. 2021. Comput. Vis. (ICCV), Seoul, South Korea, Oct. 2019, pp. 8419–8428.
[15] J. Lu, K. N. Plataniotis, and A. N. Venetsanopoulos, “Regularized [41] H. Yao et al., “Graph few-shot learning via knowledge transfer,” in
discriminant analysis for the small sample size problem in face recogni- Proc. 34th AAAI Conf. Artif. Intell., New York, NY, USA, 2020,
tion,” Pattern Recognit. Lett., vol. 24, no. 16, pp. 3079–3087, Dec. 2003. pp. 6656–6663.
[16] C. Lemke, M. Budka, and B. Gabrys, “Metalearning: A survey of [42] V. G. Satorras and J. B. Estrach, “Few-shot learning with graph neural
trends and technologies,” Artif. Intell. Rev., vol. 44, no. 1, pp. 117–130, networks,” in Proc. 6th Int. Conf. Learn. Represent. (ICLR), Vancouver,
Jun. 2015. BC, Canada, 2018, pp. 1–13.

zed licensed use limited to: MINISTERE DE L'ENSEIGNEMENT SUPERIEUR ET DE LA RECHERCHE SCIENTIFIQUE. Downloaded on April 24,2023 at 16:15:31 UTC from IEEE Xplore. Restriction
 5009812 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 70, 2021

[43] C. Jiang, H. Xu, X. Liang, and L. Lin, “Hybrid knowledge routed Huikai Shao (Graduate Student Member, IEEE)
modules for large-scale object detection,” in Proc. Adv. Neural Inf. received the B.Sc. degree from Chongqing Univer-
Process. Syst., Montreal, QC, Canada 2018, pp. 1552–1563. sity, Chongqing, China, in 2017. He is currently pur-
[44] Y. Liu et al., “Learning to propagate labels: Transductive propagation suing the Ph.D. degree with the School of Automa-
network for few-shot learning,” in Proc. 7th Int. Conf. Learn. Repre- tion Science and Engineering, Xi’an Jiaotong Uni-
sent. (ICLR), New Orleans, LA, USA, 2019, pp. 1–14. versity, Xi’an, China.
[45] Y. Wang, C. Xu, C. Liu, L. Zhang, and Y. Fu, “Instance credibility infer- His main research interests are biometrics and
ence for few-shot learning,” in Proc. IEEE/CVF Conf. Comput. Vis. Pat- computer vision.
tern Recognit. (CVPR), Seattle, WA, USA, Jun. 2020, pp. 12833–12842.
[46] S. Ravi and A. Beatson, “Amortized Bayesian meta-learning,” in Proc.
7th Int. Conf. Learn. Represent. (ICLR), New Orleans, LA, USA, 2019,
pp. 1–14.
[47] H. Li, W. Dong, X. Mei, C. Ma, F. Huang, and B. Hu, “LGM-Net:
Learning to generate matching networks for few-shot learning,” in Proc. Dexing Zhong (Member, IEEE) received the B.Sc.
36th Int. Conf. Mach. Learn. (ICML), Long Beach, CA, USA, 2019, and Ph.D. degrees from Xi’an Jiaotong University,
pp. 3825–3834. Xi’an, China, in 2005 and 2010, respectively.
[48] G. Koch, R. Zemel, and R. Salakhutdinov, “Siamese neural networks He was a Visiting Scholar with the University
for one-shot image recognition,” in Proc. ICML Deep Learn. Workshop, of Illinois at Urbana–Champaign, Champaign, IL,
Lille, France, 2015, pp. 1–30. USA. He is currently an Associate Professor with
[49] O. Vinyals, C. Blundell, T. Lillicrap, K. Kavukcuoglu, and D. Wierstra, the School of Automation Science and Engineering,
“Matching networks for one shot learning,” in Proc. Adv. Neural Inf. Xi’an Jiaotong University. His main research inter-
Process. Syst., Barcelona, Spain, 2016, pp. 3630–3638. ests are biometrics and computer vision.
[50] E. Ustinova and V. S. Lempitsky, “Learning deep embeddings with
histogram loss,” in Proc. Adv. Neural Inf. Process. Syst., Barcelona,
Spain, 2016, pp. 4170–4178.
[51] D. Zhang, Z. Guo, G. Lu, L. Zhang, and W. Zuo, “An online system
of multispectral palmprint verification,” IEEE Trans. Instrum. Meas., Xuefeng Du received the B.Eng. degree from the
vol. 59, no. 2, pp. 480–490, Feb. 2010. School of Automation Science and Engineering,
[52] H. Shao, D. Zhong, and X. Du, “Efficient deep palmprint recognition via Xi’an Jiaotong University, Xi’an, China, in 2020.
distilled hashing coding,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern He is currently pursuing the Ph.D. degree (major in
Recognit. Workshops (CVPRW), Long Beach, CA, USA, Jun. 2019, computer science) with the University of Wisconsin–
pp. 714–723. Madison, Madison, WI, USA.
[53] H. Shao, D. Zhong, and X. Du, “Deep distillation hashing for uncon- His main research interests are computer vision
strained palmprint recognition,” IEEE Trans. Instrum. Meas., vol. 70, and deep learning.
pp. 1–13, 2021, doi: 10.1109/TIM.2021.3053991.
[54] A.-S. Ungureanu, S. Thavalengal, T. E. Cognard, C. Costache, and
P. Corcoran, “Unconstrained palmprint as a smartphone biometric,”
IEEE Trans. Consum. Electron., vol. 63, no. 3, pp. 334–342, Aug. 2017.
[55] Z. Sun, T. Tan, Y. Wang, and S. Li, “Ordinal palmprint represention
for personal identification [represention read representation],” in Proc. Shaoyi Du (Member, IEEE) received the dual bach-
IEEE Comput. Soc. Conf. Comput. Vis. Pattern Recognit., Orlando, FL, elor’s degrees in computational mathematics and
USA, Jun. 2005, pp. 279–284. computer science, the M.S. degree in applied math-
[56] A. Kumar and S. Shekhar, “Personal identification using multibiometrics ematics, and the Ph.D. degree in pattern recogni-
rank-level fusion,” IEEE Trans. Syst., Man, Cybern., C (Appl. Rev.), tion and intelligence systems from Xi’an Jiaotong
vol. 41, no. 5, pp. 743–752, Sep. 2011. University, Xi’an, China, in 2002, 2005, and 2009,
[57] D. Zhong, H. Shao, and X. Du, “A hand-based multi-biometrics via respectively.
deep hashing network and biometric graph matching,” IEEE Trans. Inf. He is currently a Professor with Xi’an Jiaotong
Forensics Security, vol. 14, no. 12, pp. 3140–3150, Dec. 2019. University. His current research interests include
[58] F. Ma, X. Zhu, C. Wang, H. Liu, and X.-Y. Jing, “Multi-orientation computer vision, machine learning, and pattern
and multi-scale features discriminant learning for palmprint recognition,” recognition.
Neurocomputing, vol. 348, pp. 169–178, Jul. 2019.
[59] L. Fei, B. Zhang, Y. Xu, D. Huang, W. Jia, and J. Wen, “Local
discriminant direction binary pattern for palmprint representation and Raymond N. J. Veldhuis (Senior Member, IEEE)
recognition,” IEEE Trans. Circuits Syst. Video Technol., vol. 30, no. 2, received the degree from the University of Twente,
pp. 468–481, Feb. 2020. Twente, The Netherlands, in 1981, and the Ph.D.
[60] T.-H. Chan, K. Jia, S. Gao, J. Lu, Z. Zeng, and Y. Ma, “PCANet: A degree from Nijmegen University, Nijmegen, The
simple deep learning baseline for image classification?” IEEE Trans. Netherlands, on a thesis entitled Adaptive Restora-
Image Process., vol. 24, no. 12, pp. 5017–5032, Dec. 2015. tion of Lost Samples in Discrete-Time Signals and
[61] H. O. Song, Y. Xiang, S. Jegelka, and S. Savarese, “Deep metric Digital Images, in 1988.
learning via lifted structured feature embedding,” in Proc. IEEE Conf. From 1982 to 1992, he was a Researcher
Comput. Vis. Pattern Recognit. (CVPR), Seattle, WA, USA, Jun. 2016, with Philips Research Laboratories, Eindhoven, The
pp. 4004–4012. Netherlands, in various areas of digital signal
[62] X. Wang, X. Han, W. Huang, D. Dong, and M. R. Scott, “Multi- processing. From 1992 to 2001, he was involved in
similarity loss with general pair weighting for deep metric learning,” the field of speech processing. He is currently a Full Professor in biometric
in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), Long pattern recognition with the University of Twente, where he is leading a
Beach, CA, USA, Jun. 2019, pp. 5017–5025. research team in this field. The main research topics are face recognition (2-D
[63] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for and 3-D), fingerprint recognition, vascular pattern recognition, multibiometric
image recognition,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. fusion, and biometric template protection. The research is both applied and
(CVPR), Seattle, WA, USA, Jun. 2016, pp. 770–778. fundamental.

zed licensed use limited to: MINISTERE DE L'ENSEIGNEMENT SUPERIEUR ET DE LA RECHERCHE SCIENTIFIQUE. Downloaded on April 24,2023 at 16:15:31 UTC from IEEE Xplore. Restriction