Finger-UNet: A U-Net Based Multi-Task Architecture for Deep

Fingerprint Enhancement

Ekta Gavas

and Anoop Namboodiri

Center for Visual Information Technology, International Institute of Information Technology, Hyderabad, India

Keywords:

Fingerprint Enhancement, Fingerprint Quality, Image Enhancement, Multi-Task Learning.

Abstract:

For decades, ﬁngerprint recognition has been prevalent for security, forensics, and other biometric appli-

cations. However, the availability of good-quality ﬁngerprints is challenging, making recognition difﬁcult.

Fingerprint images might be degraded with a poor ridge structure and noisy or less contrasting backgrounds.

Hence, ﬁngerprint enhancement plays a vital role in the early stages of the ﬁngerprint recognition/veriﬁcation

pipeline. In this paper, we investigate and improvise the encoder-decoder style architecture and suggest intu-

itive modiﬁcations to U-Net to enhance low-quality ﬁngerprints effectively. We investigate the use of Discrete

Wavelet Transform (DWT) for ﬁngerprint enhancement and use a wavelet attention module instead of max

pooling which proves advantageous for our task. Moreover, we replace regular convolutions with depthwise

separable convolutions, which signiﬁcantly reduces the memory footprint of the model without degrading the

performance. We also demonstrate that incorporating domain knowledge with ﬁngerprint minutiae prediction

task can improve ﬁngerprint reconstruction through multi-task learning. Furthermore, we also integrate the

orientation estimation task to propagate the knowledge of ridge orientations to enhance the performance fur-

ther. We present the experimental results and evaluate our model on FVC 2002 and NIST SD302 databases to

show the effectiveness of our approach compared to previous works.

1 INTRODUCTION

Fingerprints are one of the most crucial biometric

traits due to their characteristic of being unique and

permanent (lifelong) to every individual (Jain et al.,

2004). In addition, they are comparatively easy to

acquire (Jain et al., 2004). So, ﬁngerprints are very

commonly used biometrics for identiﬁcation. Also,

the recognition of ﬁngerprints collected from crime

scenes is helpful in criminal investigations and foren-

sic applications. However, the acquisition of ’good’

quality ﬁngerprints is not trivial. A ﬁngerprint is iden-

tiﬁed based on its unique ridge-valley structure with a

well-deﬁned frequency, orientation, and the location

of special points called minutia. In reality, a ﬁnger-

print may be degraded due to various reasons like sen-

sor defects, noise, oily skin, ﬁnger cuts/wounds, or

uneven pressure while acquiring. It may have poor

ridge structure, overlapping backgrounds, and low

contrast due to collection from crime scenes. Due

to these factors, the performance of ﬁngerprint sys-

https://orcid.org/0000-0001-6437-3357

https://orcid.org/0000-0002-4638-0833

tems gets affected. Fingerprint enhancement deals

with this issue by uplifting the quality of ﬁngerprints

to, at the least, recover the ﬁngerprint structure to

the maximum extent possible. Whether it is match-

ing or recognition, ﬁngerprint enhancement thus be-

comes an essential step in pre-processing in cases

where good-quality prints are rare.

Traditionally, ﬁltering techniques and other clas-

sical image processing methods were used to improve

the ridge clarity, and ﬁngerprint quality (Chikkerur

et al., 2005; Greenberg et al., 2002; Hong et al., 1998;

Kim et al., 2002; Yang et al., 2002). Then, with the

advent of deep learning, neural networks, particularly

convolutional neural networks (CNNs), are being em-

ployed to tackle this problem (Li et al., 2018; Qian

et al., 2019; Joshi et al., 2019). In this paper, to re-

cover good quality ﬁngerprints, we explore a popular

architecture, U-Net (Ronneberger et al., 2015), essen-

tially an encoder-decoder-style architecture with skip

connections. This paper focuses on improving the ba-

sic U-Net architecture in order to improve ﬁngerprint

quality in a robust and intuitive manner. We suggest

that these changes improve the ﬁngerprint quality in

addition to reducing the network parameters.

Gavas, E. and Namboodiri, A.

Finger-UNet: A U-Net Based Multi-Task Architecture for Deep Fingerprint Enhancement.

DOI: 10.5220/0011687400003417

In Proceedings of the 18th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2023) - Volume 4: VISAPP, pages

309-316

ISBN: 978-989-758-634-7; ISSN: 2184-4321

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

309

Research studies show that CNNs, not being

noise-robust, noise gets enlarged as data propagates

through the layers of CNNs after several epochs of

training which may signiﬁcantly impact learning and

can even lead to overﬁtting (Xie et al., 2019; Geirhos

et al., 2018; Zhao et al., 2022; Li et al., 2021; Liu

et al., 2019). The downsampling operation of CNNs

is responsible for weak noise-robustness and loss of

information. In the frequency domain, DWT can pro-

vide high-quality downsampling, signiﬁcantly reduc-

ing this information loss. It decomposes the 2D im-

age into four frequency components, and the noise-

containing component is ﬁltered and dropped; hence

it is not forwarded into the network layers, avoiding

noise propagation. In this paper, we use a Wavelet-

Attention block proposed in (Zhao et al., 2022) in our

U-Net architecture as a downsampling layer, which

constrains the noise from high-frequency components

obtained with DWT to propagate further, whereas

information in low-frequency components is unaf-

fected.

Further, we incorporate domain knowledge into

our learning for the model to better understand the ﬁn-

gerprint structure. We achieve this by adding minutia-

prediction and orientation estimation branches with a

multi-task learning approach, which further helps to

improve performance. Moreover, we replace standard

convolution layers with depthwise separable convolu-

tions (Chollet, 2017), which does not impact the per-

formance considerably but helps to reduce the model

parameters drastically.

1.1 Related Work

In past decades, several works have been proposed

for ﬁngerprint enhancement that involves applying

traditional or classical image processing techniques.

The research focused on improving the ridge struc-

ture and increasing the contrast in images (Green-

berg et al., 2002). It started with most basic tech-

niques like histogram equalization (Ezhilmaran and

Adhiyaman, 2014) to using Fourier transforms (Sher-

lock et al., 1992; Chikkerur et al., 2005; Rahman

et al., 2008) to remove noise from images. (Chikkerur

et al., 2005) extended the use of short-term Fourier

Transform (STFT) analysis to 2D ﬁngerprint images

and estimated the intrinsic properties of ﬁngerprints.

Gabor ﬁltering was used to estimate the orientation

and frequency ﬁelds which were later used to enhance

ﬁngerprints (Hong et al., 1998; Kim et al., 2002; Yang

et al., 2002). (Liu et al., 2014) proposed a dictionary-

based approach where dictionaries were created with

a set of Gabor ﬁlters, and the multi-scale represen-

tation is iteratively applied to recover the enhanced

image. (Feng et al., 2012) proposed a path-based dic-

tionary approach for orientation estimation.

For many years, much research has been carried

out on neural networks, especially CNNs. With this,

many past works proposed deep architectures for ﬁn-

gerprint enhancement. (Qian et al., 2019) introduced

a deep network with dense blocks called DenseUNET

to improve image quality in pixel-to-pixel and end-to-

end manner. (Joshi et al., 2019) incorporated adver-

sarial training using GANs for ﬁngerprint enhance-

ment. In another interesting work, (V and Sivaswamy,

2018) posed the ﬁngerprint denoising problem as a

segmentation (task) using M-net-based architecture.

Few previous works had considered adding orienta-

tion knowledge to guide the enhancement task. (Li

et al., 2018) proposed a deep architecture called Fin-

gerNet with enhancement and orientation deconvo-

lution branches to enhance images using multi-task

learning in two-stage training. They posed coarse ori-

entation estimation as a classiﬁcation problem with a

quantized orientation ﬁeld.

Generally, CNNs work with the matrix of pixels in

the spatial domain. In contrast, the frequency domain

deals with how these pixel values change in the spa-

tial domain. Several mathematical transforms exist in

the frequency domain, including Fourier, Laplace, Z,

and wavelet transform. Wavelet transform is widely

used in signal processing applications, image denois-

ing (Kimlyk and Umnyashkin, 2018; Ismael et al.,

2016), and compression (Chowdhury and Khatun,

2012; Kanagaraj and Muneeswaran, 2020). It decom-

poses the image information into signal details and

approximations, commonly known as high-frequency

and low-frequency components. Previous studies

show several attempts have been made to incorpo-

rate DWT into CNNs (Bae et al., 2017; Liu et al.,

2019; Li et al., 2021; Zhao et al., 2022). (Bae et al.,

2017) showed that CNNs could beneﬁt from learning

about wavelet subbands and proposed a wavelet resid-

ual network (WavResNet). (Duan et al., 2017) applied

dual-tree complex wavelet transform (DT-CWT) and

designed a Convolutional-Wavelet Neural Network

(CWNN) to suppress noise and extract features ro-

bustly from SAR images. (Liu et al., 2019) proposed a

multi-level wavelet CNN (MWCNN) model which in-

tegrates wavelet transform into CNN to reduce feature

map resolution and increase receptive ﬁeld. (Li et al.,

2021) designed DWT/IDWT layer for integration into

deep networks. Later, (Zhao et al., 2022) modiﬁed

these layers to include a wavelet attention module to

retain detailed information in high-frequency com-

ponents in DWT. In the domain of ﬁngerprint bio-

metrics also, research has been carried out to show

the effectiveness of wavelet transform for ﬁngerprint

VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications

310

enhancement task (Zhang et al., 2002; Hsieh et al.,

2003). Hence, this paper combines the advantages

of deep networks over traditional ﬁltering techniques

and wavelet attention module to design an architec-

ture for ﬁngerprint enhancement.

1.2 Contributions

We made the following contributions to this paper:

1. We propose a U-Net-based architecture with the

DWT-based wavelet attention block for ﬁnger-

print enhancement.

2. We demonstrate that using minutia detection and

orientation estimation branches in a multi-task

manner can guide enhancement task to use do-

main knowledge to improve performance further.

3. We evaluate our model with publicly available

datasets FVC 2002 and NIST SD302, and show

that our model performs well in the context

of structural similarity, ﬁngerprint quality, and

matching.

2 METHODOLOGY

The ﬁngerprint enhancement task can be related to

image denoising. However, we need to consider the

inherent properties of biometric data and that it should

be handled differently than normal real-world im-

ages. Autoencoder-style architectures are pretty pop-

ular for image-denoising tasks. We design an archi-

tecture called FingerUNET, based on the popular U-

Net (Ronneberger et al., 2015) originally proposed

for biomedical image segmentation tasks but has been

shown to perform well on ﬁngerprint enhancement

task (Qian et al., 2019; Liu and Qian, 2020). In this

paper, we offer a modiﬁed U-Net targeted to handle

ﬁngerprint data for enhancement task.

2.1 Wavelet Transform as Pooling

Wavelet transform can be said as a transformation

that maps the signal to a multi-resolution representa-

tion. Wavelet has been combined with neural network

for function approximation (Zhang and Benveniste,

1992), signal representation and classiﬁcation (Szu

et al., 1992). (Li et al., 2021) proposed DWT layers

that decompose a 2D image into its frequency com-

ponents. Here, the pooling is performed with down-

sampling operation instead of max pooling or average

pooling to avoid information loss and aliasing effect.

It also increases the noise robustness of CNNs.

Given 2D data X, the DWT usually does 1D DWT

on every row and column, resulting in four frequency

components X

, X

, and X

. X

is the low-

frequency component of input X, representing the

main information, including the basic structure in the

image; X

, X

, and X

are three high-frequency

components that save the horizontal, vertical, and di-

agonal details of X, respectively. (Li et al., 2021) de-

signed these DWT/IDWT layers in Pytorch and made

DWT/IDWT operations differentiable and compatible

with CNNs.

= LXL

, X

= HXL

, X

= HXH

(1)

where matrix L and H are the cyclic matrix composed

of wavelet low-pass ﬁlter {l

}

k∈Z

and high-pass ﬁl-

ter {h

}

k∈Z

respectively. Here L and H are as in (Li

et al., 2021). In FingerUNET, we make use of wavelet

attention block (WA Block) proposed in (Zhao et al.,

2022) built by modifying the above DWT layer. The

wavelet attention block can be deﬁned as

= σ( f (X

, X

)) (2)

= X

∗ x

(3)

Z = f (X

, A

) (4)

where f represents the feature aggregation function, σ

denotes the softmax function. Once the four compo-

nents are obtained using DWT, the WA block takes the

horizontal feature X

and vertical feature X

and

aggregates them by element-wise addition as a global

detail feature. Then this feature is normalized using

the softmax function. The normalized feature x

and

the low-frequency component X

are used to gener-

ate the attention map A

through element-wise multi-

plication. Finally, the original low-frequency compo-

nent X

is added to the attention map A

by element-

wise operation and given as output Z. As the compo-

nent X

does not contain additional information for

ﬁngerprint ridges, it is not used in WA block. We

replace the max-pooling layers in vanilla U-Net with

this WA block and notice an improvement in the ﬁn-

gerprint enhancement task. IDWT layer is used to re-

construct the output back to spatial domain (Li et al.,

2021).

2.2 Reconstruction Loss

The choice of the loss function is crucial in any neural

network training. Earlier works (Burger et al., 2012;

Dong et al., 2014) used mean squared error (MSE)

or l

loss for reconstruction in image enhancement

tasks. (Zhao et al., 2016) pointed out several limi-

tations to using l

for image restoration tasks. l

does

not correlate well with human perception of image

Finger-UNet: A U-Net Based Multi-Task Architecture for Deep Fingerprint Enhancement

311

Figure 1: Architecture of our proposed approach with enhancement, minutia detection and orientation estimation branches in

a multi-task learning setting.

quality (Zhang et al., 2012), due to the assumption

that the impact of noise is independent of the local

characteristics of the image. Moreover, l

penalizes

larger errors but tolerates small ones without consid-

ering the underlying structure of the image. In con-

trast, the human visual system (HVS) is more sensi-

tive to luminance and color variations in texture-less

regions (Zhao et al., 2016; Winkler and Susstrunk,

2004). (Zhao et al., 2016) suggested using l

loss in-

stead of l

as it does not over-penalize large errors and

it also helps reduce artifacts introduced by l

. We uti-

lize the effectiveness of using l

as a reconstruction

loss for ﬁngerprint enhancement. Even though the

images are grayscale, we witness a performance im-

provement compared to the l

counterpart, as demon-

strated in later sections.

2.3 Minutia Detection and Orientation

Estimation

Every ﬁngerprint has a well-deﬁned ridge-valley

structure and a frequency associated with it. The

presence of speciﬁc patterns like whorl, loop, and

arch in a ﬁngerprint at different locations makes it

unique. Ridge endings and bifurcations (minutiae)

are important in ﬁngerprint matching. Hence, ﬁnger-

print images prove to be very different from normal

real-world images and hence they need to be dealt

with differently. These properties of ﬁngerprint data

should not be neglected in learning. For this, adding

domain knowledge to our network training becomes

useful to model the data better. Moreover, in the case

of low-quality images where the structure is not intu-

itive, the knowledge of ﬁngerprint properties can help

the network to learn and predict the structure better.

Previous works utilize ridge orientations when en-

hancing ﬁngerprints (Hong et al., 1998; Kim et al.,

2002; Yang et al., 2002; Li et al., 2018). Orientation

estimation gives the direction of the gradient of the

ﬁngerprint segment which can be important in ﬁnger-

print enhancement. (Li et al., 2018) trained enhance-

ment and orientation estimation branches in an end-

to-end manner and demonstrated the effectiveness of

orientation knowledge to aid the enhancement. The

orientation estimation branch in our work predicts

vectorized orientation ﬁelds, speciﬁcally sine and co-

sine values. This eliminates the need to segregate ori-

entation patches into 20 ﬁxed sets of classes (Li et al.,

2018) and allows for the estimation of precise ridge

orientation.

We also incorporate the minutia detection branch

for using domain knowledge. (Darlow and Ros-

man, 2017) proposed a minutia extraction network

called MENet, which outputs the minutiae probabil-

ities map, which is later processed to determine pre-

cise minutia locations. To propagate the informa-

tion about ﬁngerprint properties to the enhancement

branch, we use a gray-scale image with minutiae loca-

tions marked with a white dot on a black background

as ground-truth as shown in the minutia branch (Task

2) image in Figure 1.

We approach the problem at hand using multi-

task learning by optimizing ﬁngerprint enhancement,

minutia detection, and orientation tasks simultane-

ously. We use a shared encoder that learns the feature

representations from noisy ﬁngerprint images and one

decoder for each of the three tasks, as shown in Figure

1. We use the l

loss mentioned in Section 2.2 for the

enhancement task denoted by L

. As we are solving

orientation estimation as a regression problem, we use

loss or MSE for this branch. We denote this loss by

. For the minutia detection branch, the values in the

VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications

312

ground-truth map are zero or one based on the pres-

ence of minutia locations, hence we use binary cross

entropy as the loss function and it is denoted by L

The ﬁnal loss L is the summation of L

, L

, and L

weighted by scalars λ

, λ

and λ

respectively.

total

= λ

+ λ

(5)

2.4 Operation Approximation with

Depthwise Separable Convolution

Standard convolution operation applies spatial and

channel interactions by multiplying values over sev-

eral spatial pixels and all the channels. The idea of

depthwise separable convolution (Chollet, 2017) is

to disentangle the two by using each ﬁlter channel

only at one input channel. Then, we use a 1x1 ﬁl-

ter to cover the depth dimension. Though this method

is an approximation to the usual convolution opera-

tion, it does not incur a considerable drop in perfor-

mance. In addition, it reduces the number of model

parameters by 30% in our experiments which in turn

helps to avoid the over-ﬁtting problem. Hence, in our

work, we replace the convolution layers with depth-

wise separable convolutions.

3 EXPERIMENTS

3.1 Dataset

The ﬁngerprint datasets available in the public do-

main are either not huge enough to train a deep net-

work or do not contain enhanced/clean impressions

for ground-truth. So, for all the experiments in this

paper, we are using synthetically generated ﬁnger-

prints from SFinGe (Cappelli et al., 2002). SFinGe

can generate ﬁngerprints with the required number

of impressions and varying ridge structures and pat-

terns. A wide range of noise and degradation like

skin elasticity, noise, pressure, scratches, etc., can

also be added. We generated 10,000 ﬁngerprint pairs

(degraded and ground-truth image pairs) for training,

1,000 for validation, and 3,000 for testing. The gener-

ated data contain varying types and degrees of avail-

able noises and backgrounds (optical, capacitive sen-

sors, and no background). Figure 2 shows sample in-

put and ground-truth pairs from the dataset in the ﬁrst

two rows. For evaluation, we use publicly available

datasets FVC 2002 (Maio et al., 2002) and NIST Spe-

cial Database (SD) 302 (Fiumara et al., 2019). FVC

2002 consists of ﬁngerprints from optical, capacitive

sensors, and synthetic ﬁngerprints available in four

Figure 2: Enhanced images (bottom row) with correspond-

ing degraded input (top row) and ground-truth (middle row)

with our approach on SFinGe synthetic test set. SSIM is

reported between each enhanced and ground-truth pair.

sets, of which 80 images are publicly available in each

set. NIST SD302 has plain, rolled and touch-free im-

pressions captured from various devices. We use the

subset 302d containing 5141 ﬁngerprint images ac-

quired from 4 different auxiliary devices.

3.2 Training

We use the PyTorch framework for all experiments in

this paper. We apply various data augmentations like

random translation, rotation, ﬂip (horizontal/vertical),

and shear, for the network to generalize well. We ap-

ply the same degree of augmentations to each input-

ground truth pair for consistency. The images are re-

sized to ﬁxed dimensions of 400x256. The hyperpa-

rameters are chosen using grid search. The model is

trained with Adam optimizer with a learning rate of

0.001. The batch size is set as 32. The loss weights

, λ

, and λ

are 0.8, 0.1, 0.1 respectively. The net-

work is trained on four GPUs in data parallel mode,

and each GPU is NVIDIA GeForce TITAN X with

16 GB RAM. As the training is end-to-end, the losses

from minutia and orientation guide the enhancement

branch for better ﬁngerprint output.

4 RESULTS AND ANALYSIS

For evaluation, we make use of standard metrics like

SSIM, RMSE, and PSNR (V and Sivaswamy, 2018).

Additionally, we utilize NFIQ2 (Tabassi et al., 2021)

package from NIST’s NBIS (Ko, 2007) to measure

the quality of ﬁngerprint images. Quality scores

can range from 1 to 100. Moreover, we present the

average matching scores of genuine pairs on all four

subsets of FVC2002 using BOZORTH3 (Ko, 2007).

SD302 does not contain multiple impressions of a

Finger-UNet: A U-Net Based Multi-Task Architecture for Deep Fingerprint Enhancement

313

(a) FVC 2002 dataset. (b) NIST SD302 dataset.

Figure 3: Illustration of enhancement results with our approach on samples from test datasets.

ﬁnger, so matching performance can not be obtained.

Ridge Structure Preservation: In Figure 2, we

show a few sample images from our SFinge test

set with corresponding ground-truth and enhanced

images. We see the SSIM values are higher which

suggests that our approach tries to preserve the ridge

structure while performing enhancement on degraded

input. In addition to this, Table 1 reports SSIM values

for various combinations of techniques suggested in

this paper. In Figure 3, we show the results of our

approach on both datasets.

Fingerprint Quality Analysis: We report the

average NFIQ2 scores on the test sets FVC 2002

and NIST SD302 in Table 2. From the results,

we say that the ﬁngerprint quality improved by a

signiﬁcant amount of 58% in the case of FVC 2002,

whereas it improved by 23% in the SD302 dataset

after enhancing the raw images with our approach.

Further, our approach gives comparable results with

the previous works. Moreover, we also present the

NFIQ2 scores on the SFinGe test set in Table 1 which

supports our approach to use WA Block, depthwise

separable convolutions, and domain knowledge.

Matching Performance: We report and com-

pare the average matching scores of genuine pairs

of our approach with raw images and previous

works in Table 3. The results suggest our approach

with the inclusion of domain knowledge from the

minutia and orientation branch is able to retain the

minutiae from the degraded images which increases

the matching score and performs well in comparison

to earlier works. This suggests the effectiveness of

our approach in the pipeline of ﬁngerprint matching.

Ablation Study: We report the evaluation met-

rics for the techniques discussed in the paper as

a part of the ablation study in Table 1. We show

Figure 4: Illustration of failure scenarios observed with our

approach. Top row represents degraded input and bottom

row denotes the enhanced images with our model.

how each discussed method in this paper helps to

improve the enhancement further. The depthwise

separable convolution do not signiﬁcantly impact

the performance but help reduce the model param-

eters to a large extent. Multi-task learning with

minutia detection and orientation branches gives

a clear performance improvement. Moreover, the

use of wavelet attention block further improves

the SSIM and NFIQ2. Overall, the combination of

these techniques results in the best performing model.

Observed Challenges: We observed that our ap-

proach Finger-UNet performs well from our experi-

mental results. However, we also observed scenar-

ios where our approach did not perform well. A few

of these cases are demonstrated in Figure 4. We saw

that in cases where the input has severe artifacts, the

model mistakes it for a portion of the ﬁngerprint and

tries to enhance it. Moreover, in cases where the input

is too dark or too light to correctly ﬁgure out the ridge

structure, the model fails to predict good ﬁngerprints

in those areas. In addition, if the input contains a nail,

it fails to discriminate the nail from the ﬁngerprint.

We believe these issues arise as the network has not

seen such data during training, as the SFinge dataset

does not contain samples with nails or artifacts.

VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications

314

Table 1: Ablation study: Evaluation performance on

SFinGe test dataset with different modiﬁcations to U-Net

suggested in this paper.

Approach SSIM RMSE PSNR NFIQ2

Raw Images 0.605 117.31 6.89 36.42

loss) 0.863 43.67 12.78 47.72

loss) 0.883 41.83 14.12 49.45

Depth. Sep. (DS) 0.890 41.25 14.17 49.78

WA Block (WA) 0.919 40.31 14.43 51.01

Minutia (M) 0.934 37.92 17.15 52.86

Orientation (O) 0.928 38.58 16.81 51.32

M+O 0.943 37.43 17.62 53.11

M+O+WA 0.954 36.91 18.47 55.32

M+O+WA+DS 0.955 36.78 18.81 55.34

Table 2: Average NFIQ2 scores of the images from FVC

2002 and NIST SD302 datasets. Higher scores represent

higher ﬁngerprint image quality.

Dataset FVC 2002 NIST SD302

Raw Images 35.10 46.97

Joshi et al. 54.11 56.84

Hong et al. 56.01 58.23

Ours 56.26 58.49

Table 3: Average matching scores from BOZORTH3 on dif-

ferent subsets (DB1, DB2, DB3 and DB4) of FVC 2002

dataset. Higher the scores, better is the approach. Feature

extraction was performed using MINDTCT.

Database DB1 DB2 DB3 DB4

Raw Images 52.77 48.62 45.21 50.26

Joshi et al. 71.34 71.06 67.08 68.20

Hong et al. 73.52 72.31 69.12 70.29

Ours 74.01 72.61 69.10 71.08

5 CONCLUSION

In this paper, we modiﬁed vanilla U-Net, combining

multiple techniques to improve the ﬁngerprint quality

with synthetic data from SFinGe. We evaluated our

model on two public ﬁngerprint datasets FVC 2002

and NIST SD302. The network is robust enough to re-

cover ﬁngerprints even with various degrees of degra-

dation. From the experimental results, we say that l

loss performed well for this task, along with domain

knowledge from minutia and orientation branches,

which improved performance above baselines. We

also discussed the challenging cases in our experi-

ments and the possible solutions. Moreover, using

the wavelet attention block helped improve the per-

formance. In future works, we plan to explore and de-

sign new quality metrics better suited for ﬁngerprint

enhancement along with other deep architectures.

REFERENCES

Bae, W., Yoo, J., and Chul Ye, J. (2017). Beyond deep

residual learning for image restoration: Persistent

homology-guided manifold simpliﬁcation. In Pro-

ceedings of the IEEE conference on computer vision

and pattern recognition workshops.

Burger, H. C., Schuler, C. J., and Harmeling, S. (2012).

Image denoising: Can plain neural networks compete

with bm3d? In 2012 IEEE conference on computer

vision and pattern recognition. IEEE.

Cappelli, R., Maio, D., and Maltoni, D. (2002). Synthetic

ﬁngerprint-database generation. In 2002 International

Conference on Pattern Recognition, volume 3. IEEE.

Chikkerur, S., Govindaraju, V., and Cartwright, A. N.

(2005). Fingerprint image enhancement using stft

analysis. In International Conference on Pat-

tern Recognition and Image Analysis, pages 20–29.

Springer.

Chollet, F. (2017). Xception: Deep learning with depthwise

separable convolutions. In Proceedings of the IEEE

conference on computer vision and pattern recogni-

tion.

Chowdhury, M. M. H. and Khatun, A. (2012). Image com-

pression using discrete wavelet transform. Interna-

tional Journal of Computer Science Issues (IJCSI),

9(4):327.

Darlow, L. N. and Rosman, B. (2017). Fingerprint minutiae

extraction using deep learning. In 2017 IEEE Interna-

tional Joint Conference on Biometrics (IJCB). IEEE.

Dong, C., Loy, C. C., He, K., and Tang, X. (2014). Learn-

ing a deep convolutional network for image super-

resolution. In European conference on computer vi-

sion. Springer.

Duan, Y., Liu, F., Jiao, L., Zhao, P., and Zhang, L. (2017).

Sar image segmentation based on convolutional-

wavelet neural network and markov random ﬁeld. Pat-

tern Recognition, 64:255–267.

Ezhilmaran, D. and Adhiyaman, M. (2014). A review study

on ﬁngerprint image enhancement techniques. Inter-

national Journal of Computer Science & Engineering

Technology (IJCSET) ISSN, pages 2229–3345.

Feng, J., Zhou, J., and Jain, A. K. (2012). Orientation ﬁeld

estimation for latent ﬁngerprint enhancement. IEEE

transactions on pattern analysis and machine intelli-

gence, 35(4):925–940.

Fiumara, G. P., Flanagan, P. A., Grantham, J. D., Ko, K.,

Marshall, K., Schwarz, M., Tabassi, E., Woodgate, B.,

Boehnen, C., et al. (2019). Nist special database 302:

Nail to nail ﬁngerprint challenge.

Geirhos, R., Rubisch, P., Michaelis, C., Bethge, M., Wich-

mann, F. A., and Brendel, W. (2018). Imagenet-

trained cnns are biased towards texture; increasing

Finger-UNet: A U-Net Based Multi-Task Architecture for Deep Fingerprint Enhancement

315

shape bias improves accuracy and robustness. arXiv

preprint arXiv:1811.12231.

Greenberg, S., Aladjem, M., and Kogan, D. (2002). Finger-

print image enhancement using ﬁltering techniques.

Real-Time Imaging, 8(3):227–236.

Hong, L., Wan, Y., and Jain, A. (1998). Fingerprint image

enhancement: Algorithm and performance evaluation.

IEEE transactions on pattern analysis and machine

intelligence, 20(8):777–789.

Hsieh, C.-T., Lai, E., and Wang, Y.-C. (2003). An effective

algorithm for ﬁngerprint image enhancement based on

wavelet transform. Pattern Recognition, 36(2):303–

312.

Ismael, S. H., Mustafa, F. M., and Ok

us, I. T. (2016).

A new approach of image denoising based on dis-

crete wavelet transform. In 2016 World Symposium on

Computer Applications & Research (WSCAR), pages

36–40. IEEE.

Jain, A. K., Ross, A., and Prabhakar, S. (2004). An intro-

duction to biometric recognition. IEEE Transactions

on circuits and systems for video technology, 14(1):4–

20.

Joshi, I., Anand, A., Vatsa, M., Singh, R., Roy, S. D., and

Kalra, P. (2019). Latent ﬁngerprint enhancement us-

ing generative adversarial networks. In 2019 IEEE

winter conference on applications of computer vision

(WACV), pages 895–903. IEEE.

Kanagaraj, H. and Muneeswaran, V. (2020). Image com-

pression using haar discrete wavelet transform. In

2020 5th International Conference on Devices, Cir-

cuits and Systems (ICDCS). IEEE.

Kim, B.-G., Kim, H.-J., and Park, D.-J. (2002). New en-

hancement algorithm for ﬁngerprint images. In 2002

International Conference on Pattern Recognition, vol-

ume 3, pages 879–882. IEEE.

Kimlyk, M. and Umnyashkin, S. (2018). Image denois-

ing using discrete wavelet transform and edge infor-

mation. In 2018 IEEE Conference of Russian Young

Researchers in Electrical and Electronic Engineering

(EIConRus). IEEE.

Ko, K. (2007). User’s guide to nist biometric image soft-

ware (nbis).

Li, J., Feng, J., and Kuo, C.-C. J. (2018). Deep convolu-

tional neural network for latent ﬁngerprint enhance-

ment. Signal Processing: Image Communication.

Li, Q., Shen, L., Guo, S., and Lai, Z. (2021). Wavecnet:

Wavelet integrated cnns to suppress aliasing effect for

noise-robust image classiﬁcation. IEEE Transactions

on Image Processing, 30:7074–7089.

Liu, M., Chen, X., and Wang, X. (2014). Latent ﬁngerprint

enhancement via multi-scale patch based sparse repre-

sentation. IEEE Transactions on Information Foren-

sics and Security, 10(1):6–15.

Liu, M. and Qian, P. (2020). Automatic segmentation and

enhancement of latent ﬁngerprints using deep nested

unets. IEEE Transactions on Information Forensics

and Security, 16:1709–1719.

Liu, P., Zhang, H., Lian, W., and Zuo, W. (2019). Multi-

level wavelet convolutional neural networks. IEEE

Access, 7:74973–74985.

Maio, D., Maltoni, D., Cappelli, R., Wayman, J. L., and

Jain, A. K. (2002). Fvc2002: Second ﬁngerprint ver-

iﬁcation competition. In 2002 International Confer-

ence on Pattern Recognition, volume 3. IEEE.

Qian, P., Li, A., and Liu, M. (2019). Latent ﬁngerprint en-

hancement based on denseunet. In 2019 International

Conference on Biometrics (ICB), pages 1–6. IEEE.

Rahman, S. M., Ahmad, M. O., and Swamy, M. (2008).

Improved image restoration using wavelet-based de-

noising and fourier-based deconvolution. In 2008 51st

Midwest Symposium on Circuits and Systems, pages

249–252. IEEE.

Ronneberger, O., Fischer, P., and Brox, T. (2015). U-net:

Convolutional networks for biomedical image seg-

mentation. In International Conference on Medical

image computing and computer-assisted intervention,

pages 234–241. Springer.

Sherlock, B., Monro, D., and Millard, K. (1992). Algorithm

for enhancing ﬁngerprint images. Electronics letters,

18(28):1720–1721.

Szu, H. H., Telfer, B. A., and Kadambe, S. L. (1992). Neu-

ral network adaptive wavelets for signal representation

and classiﬁcation. Optical Engineering, 31(9):1907–

1916.

Tabassi, E., Olsen, M., Bausinger, O., Busch, C., Figlarz,

A., Fiumara, G., Henniger, O., Merkle, J., Ruhland,

T., Schiel, C., and Schwaiger, M. (2021). Nist ﬁnger-

print image quality 2.

V, S. A. and Sivaswamy, J. (2018). Fpd-m-net: Fin-

gerprint image denoising and inpainting using m-

net based convolutional neural networks. CoRR,

abs/1812.10191.

Winkler, S. and Susstrunk, S. (2004). Visibility of noise

in natural images. In Human Vision and Electronic

Imaging IX, volume 5292, pages 121–129. SPIE.

Xie, C., Wu, Y., Maaten, L. v. d., Yuille, A. L., and He, K.

(2019). Feature denoising for improving adversarial

robustness. In Proceedings of the IEEE/CVF confer-

ence on computer vision and pattern recognition.

Yang, J., Liu, L., and Jiang, T. (2002). Improved method

for extraction of ﬁngerprint features. In Second Inter-

national Conference on Image and Graphics, volume

4875, pages 552–558. SPIE.

Zhang, L., Zhang, L., Mou, X., and Zhang, D. (2012).

A comprehensive evaluation of full reference image

quality assessment algorithms. In 2012 19th IEEE In-

ternational Conference on Image Processing. IEEE.

Zhang, Q. and Benveniste, A. (1992). Wavelet networks.

IEEE transactions on Neural Networks, 3(6).

Zhang, W.-P., Wang, Q.-R., and Tang, Y. Y. (2002). A

wavelet-based method for ﬁngerprint image enhance-

ment. In Proceedings. International Conference on

Machine Learning and Cybernetics, volume 4. IEEE.

Zhao, H., Gallo, O., Frosio, I., and Kautz, J. (2016).

Loss functions for image restoration with neural net-

works. IEEE Transactions on computational imaging,

3(1):47–57.

Zhao, X., Huang, P., and Shu, X. (2022). Wavelet-attention

cnn for image classiﬁcation. Multimedia Systems,

28(3):915–924.

VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications

316