Generative Convolutional Networks for Latent Fingerprint Reconstruction

Jan Svoboda 1, Federico Monti 1 and Michael M. Bronstein 1, 2

1Institute of Computational Science, USI Lugano, Switzerland2Perceptual Computing, Intel, Israel

Email: {jan.svoboda; federico.monti; michael.bronstein}@usi.ch

Abstract

Performance of fingerprint recognition depends heavilyon the extraction of minutiae points. Enhancement of thefingerprint ridge pattern is thus an essential pre-processingstep that noticeably reduces false positive and negative de-tection rates. A particularly challenging setting is when thefingerprint images are corrupted or partially missing. Inthis work, we apply generative convolutional networks todenoise visible minutiae and predict the missing parts ofthe ridge pattern. The proposed enhancement approach istested as a pre-processing step in combination with severalstandard feature extraction methods such as MINDTCT,followed by biometric comparison using MCC and BO-ZORTH3. We evaluate our method on several publiclyavailable latent fingerprint datasets captured using differ-ent sensors.

1. IntroductionFingerprints have been in used as a means of person au-

thentication since ancient time. The first use of fingerprintsin a criminal investigation dates back to 1880’s, when Fran-cis Galton devised the first method for classifying finger-prints. In 1892, the first successful fingerprint-based identi-fication helped convict a murderer. Since then, fingerprint-ing has underwent massive development and is nowadaysan important part of crime scene investigation as well asmodern biometric identification systems.

Based on their acquisition process, fingerprints can becategorized as inked, sensor-scan (e.g. optical, capacitive,etc.), or latent. The first two categories are typically furthersubdivided into rolled (nail-to-nail fingerprints), flat (singlefinger) or slap (four finger flat). Those have been heavilyresearched in the past yielding several state-of-the-art meth-ods [20, 19, 22].

On the other hand, latent fingerprints are impressions ofthe papillary lines that are unintentionally left by a subject atcrime scenes. They are typically partial, blured, noisy, andexhibit poor ridge quality, as opposed to inked or sensor-

(a) Synthetic Data

(b) Real Data

Figure 1. Reconstruction of damaged or latent fingerprints usingautoencoder networks. (a) Synthetic data example. From left toright: (input, reconstruction, target); (b) Real data - a pair (input,reconstruction).

scan fingerprints. Such fingerprints are usually lifted fromthe surface by means of special chemical procedures andphotographed using high-resolution camera for further pro-cessing. Latent fingerprint matching is a challenging prob-lem and state-of-the-art methods developed for inked andsensor-scan fingerprints do not work well (they typically failto detect the minutiae or detect many false minutiae due tothe imperfections mentioned above).

In practice, latent fingerprints are analyzed with the helpof forensic examiners who perform a manual latent fin-gerprint identification procedure called ACE-V (Analysis,

arX

iv:1

705.

0170

7v1

[cs

.CV

] 4

May

201

7

Comparison, Evaluation, and Verification) [3]. This processis however very demanding and time consuming. In orderto make identification efficient, forensic experts tend to re-strict the population against which compare the latent fin-gerprints (focusing, for instance, only on suspects selectedby witnesses or other evidence). This obviously reduces thelikelihood of effectively identifying the culprit, thus makingthe overall process less reliable.

Nowadays, several systems supporting the work offorensic examiners are available. The biggest one, AFIS(Automated Fingerprint Identification System), allows ex-aminers to match latent fingerprints against large databasesin a semi-automatic manner. The process usually consists ofmanually marking the minutiae points; launching the AFISmatcher; and visually verifying the top candidate finger-prints. Such a process requires considerable amount of te-dious manual labour.

In this paper, we improve latent fingerprint recognitionby enhancing the input fingerprint image that allows us-ing standard feature extraction methods more reliably. In-spired by the previous success of convolutional autoen-coders (CAE) for image processing tasks [24, 35, 23, 28,13, 9], we design a convolutional autoencoder neural net-work capable at reconstructing high-quality fingerprint im-ages from latent fingerprints. The neural network is trainedon a synthetic dataset consisting of partial and blurry finger-print impressions containing background noise; the outputof the network is compared to the groundtruth fingerprintimage. The training set was generated using the open sourceimplementation [2] of the SFinGE fingerprint generator [8].

We evaluate our method on two publicly available la-tent fingerprint datasets: IIIT-Delhi latent fingerprint [30](for latent-to-latent fingerprints matching) and IIIT-DelhiMOLF [31] (for latent-to-sensor-scan fingerprints match-ing). We show the broad applicability of our approachand discuss the possible future directions. All the eval-uations were done using only open source software. Weused MINDTCT [27] and an implementation of [1] for fea-tures extraction, and BOZORTH3 [27] and MCC (MinutiaeCylinder Code) [6, 7, 11, 12] for fingerprints matching. Weshow that using our fingerprint enhancement method, theperformance of non-commercial fingerprint matching algo-rithms can be improved and becomes comparable to someof the commercial ones.

The rest of the paper is organized as follows. In SectionII, previous works are reviewed. Section III describes theobjective function used for training the model. Section IVdescribes the convolutional autoencoder architecture. Sec-tion V presents the experimental results. Finally, Section VIconcludes the paper and discusses possible future researchdirections.

2. Related workDifferent solutions for improving latent fingerprint

matching systems have been explored in the past. Manyworks have improved latent fingerprint matching by usingextended features that require manual annotation [16, 33].However, performing the annotation in low-quality latentfingerprints may be very time-consuming and even infeasi-ble in some settings. Other works have strived towards re-ducing the amount of manual input to only selection regionof interest (ROI) and singular points [33, 34]. Approachesperforming fusion of multiple matchers or multiple latentfingerprints were explored in [30].

A different direction for improving fingerprint match-ing concerns the enhancement of the acquired latent finger-print images. Yoon et al. [33] proposed a semi-automatedmethod for enhancing the ridge information using the es-timated orientation image. A more robust orientation fieldestimation technique based on Short-Time Fourier Trans-form and RANSAC has been proposed in [34]. Feng et al.[10] proposed an approach capable of using prior knowl-edge on the fingerprint ridge structure employing a dictio-nary of reference orientated patches. Cao et al. [5] intro-duced a coarse-to-fine dictionary-based ridge enhancementtechnique.

Most recent work similar to the proposed approach isSchuch et al. [32], who used a convolutional autoencoderfor inked and sensor-scan fingerprints ridge enhancement.Their network had a rather simple architecture and was notevaluated in challenging latent fingerprint recognition set-tings.

3. Gradient-based fingerprint similaritySince our method is based on Convolutional Neural Net-

works (CNN), the gradient analysis of the fingerprint imageridge pattern becomes a natural choice as the computationof the image gradient can be implemented using the convo-lutional operator.

Gradient filters. We compute the directional derivativesof the fingerprint image I by convolving it with the direc-tional kernel Sθ

Gθ(I) = I ? Sθ, (1)

where θ is the direction in which we want to compute thegradient (we use 0, 45, 90, and 135 degrees).

As a criterion of similarity of two fingerprint images (tar-get image It and reconstructed image Ir), we used the aver-age Mean Squared Error (MSE) on all the directions,

Egrad(It, Ir) =

∑θ∈T

1n‖(It − Ir) ? Sθ‖

22

|T |, (2)

where T = {0, 45, 90, 135} is the set of considered orien-tations and n is the number of image pixels.

Ridge pattern orientation. Orientation of the ridge pat-tern can be defined through image moments using the al-ready computed image gradients. Considering Gx = G0

and Gy = G90, we define the covariance matrix of theimage using second order central moments (µ′20, µ′02, µ′11)[14]:

Cov(I(x, y)) =

(µ′20 µ′11

µ′11 µ′02

)=

(Gxx GxyGxy Gyy

), (3)

where Gxx = gΣs? (Gx ·Gx), Gyy = gΣs

? (Gy ·Gy) andGxy = gΣs

?(Gx ·Gy). gΣ represents a Gaussian smoothingkernel with covariance Σ and · denotes the element-wiseproduct.

Eigenvectors of the covariance matrix Cov(I(x, y))point in the directions of the major and minor intensity ofimage I . This information is enough to compute the orien-tation as the angle between the eigenvector correspondingto the largest eigenvalue and the x-axis is expressed by for-mula:

Θ =1

2tan−1

(2Gxy

Gxx −Gyy

). (4)

In order to strengthen the similarity between the recon-structed image and the associated ground-truth, we furthercalculate the reliability orientation field [17]. The ridge ori-entation image is converted into a continuous vector fieldas

(Φx,Φy) = (cos(2Θ), sin(2Θ)), (5)

Subsequently, we apply Gaussian low-pass filter on the re-sulting vector field,

(Φ′x,Φ′y) = (gΣo ? Φx, gΣo ? Φy). (6)

The reliability measure R is then defined by means of min-imum inertia Imin and maximum inertia Imax as follows:

Imin =(Gyy +Gxx)− (Gxx −Gyy)Φ′x −GxyΦ′y

2, (7)

Imax = Gyy +Gxx − Imin, (8)

R = 1− IminImax

. (9)

The intuition behind is that when the ratio Imin/Imax closeto 1, there is very little orientation information at that point.

Finally, we define the orientation energy Eori and reli-ability energy Erel as the MSE of the orientation and reli-ability measures computed on the target image It and thereconstructed image Ir,

Eori =1

n‖Θ(It)−Θ(Ir)‖22, (10)

Erel =1

n‖R(It)−R(Ir)‖22. (11)

4. Fingerprint reconstruction modelFollowing the principles described in the previous sec-

tions and the guidelines presented in [29], we approachedthe fingerprint reconstruction problem using a fully convo-lutional autencoder network.

Architecture. The encoding part corresponds to a fullyconvolutional neural network comprising five convolu-tional layers. It receives as input tensors of dimensions1@260x320 and produces as output tensors of dimensions1024@8x10 (F@WxH is a compact notation for tensorswith width W, height H and F different feature channels).The convolutional layers have stride equal to 2 and areequipped with REctified Linear Unit (RELU) [25]. Batchnormalization is applied to each layer for faster convergence[15].

The output of the encoder is directly fed as an input tothe decoding part. The decoder copies the architecture ofthe encoder performing the following changes. Each con-volutional layer is replaced with de-convolutional layer (i.e.a fractionally-strided convolutional layer with stride equalto 0.5 [29]), and each RELU with a Leaky-RELU [21].In order to reproduce the original grey scale image, a fur-ther convolutional layer equipped with a sigmoid activationfunction is applied at the end of the decoder. Figure 2 de-picts our architecture.

Training objective. Based on the analysis described inSection 3, we define an objective function that allows to ef-ficiently train our convolutional autoencoder to reconstructcorrupted parts of the fingerprint images. The objectivefunction corresponds to a weighted average of three differ-ent losses:

E = Egrad + λ(Eori + Erel), (12)

where Egrad, Eori and Erel are as defined in the previoussection, and λ is a parameter weighting the contribution ofthe orientation and reliability regularizers.

Training. For the purpose of training our model, we gen-erated a dataset of synthetic fingerprints. From these fin-gerprint images, we simulated latent fingerprint images byapplying rotation, translation, directional blur and morpho-logical dilation, and blend the results with several differentbackgrounds. This way, we aim to simulate the latent fin-gerprint formation as reliably as possible. We binarize thegroudtruth (target) images using approach based on [4]. Thenetwork is applied to the synthetic latent fingerprint imagesand tries to reconstruct the underlying groundtruth image byminimizing the above objective function between the outputand the groundtruth image. Since the groundtruth images

Input1@256x32013x13 kernel

Feature maps128@128x16011x11 kernel

Feature maps256@64x809x9 kernel

Feature maps256@32x407x7 kernel

Feature maps512@16x205x5 kernel

Feature maps1024@8x105x5 kernel

Feature maps1024@16x207x7 kernel

Feature maps512@32x409x9 kernel

Feature maps256@64x8011x11 kernel

Feature maps256@128x16013x13 kernel

Feature maps128@256x3205x5 kernel

Output1@256x320 kernel

Figure 2. Our fully convolutional autoencoder.

are binarized, our model learns this way not only to recon-struct the fingerprints, but also to directly produce binaryimages.

Overall, we used 15000 synthetic fingerprint images fortraining. Data augmentation was performed by adding ran-dom Gaussian i.i.d. noise with µ = 0 and σ = 3.5 · 10−3.The network was trained for 400 epochs (each epoch con-sisting of 64 iterations). In each iteration, we fed the net-work with a batch of 12 latent/groundtruth image pairs. Weemployed Adam [18] updates with β1 = 0.5 and weight de-cay using L2 regularization with µ = 10−4. The learningrate is set to 2 · 10−4. The weighting parameter λ was set to0.1.

5. ExperimentsWe demonstrate the proposed method on several settings

of latent fingerprint recognition. The experiments were car-ried out using publicly available datasets. Latent fingerprintenhancement was evaluated using the IIIT-Delhi Latent fin-gerprint [30] and IIIT-Delhi MOLF [31] datasets.

For the first evaluation, we applies standard fingerprintrecognition algorithms on original fingerprint images andthose enhanced by our method. We employed two differ-ent feature extraction methods: ABR11 proposed by Abra-ham et al. [1] and MINDTCT from NBIS [27]. Extractedfeatures were subsequently compared using two differentmethods, BOZORTH3 [27] (abbreviated to BOZ in thefollowing) and Minutia Cylinder Code (MCC) [6, 7, 11,12]. For the other evaluations, we used the combinationABR11 + MCC as the best performing one for latent finger-print matching. Results of latent fingerprint enhancementwere presented using TopX measure and Cumulative MatchCharacteristic (CMC) curves.

Latent-to-Latent matching. We evaluated latent-tolatentfingerprint matching using the protocol described in [30].The whole IIIT-Delhi latent fingerprint dataset contained1046 samples of all the ten fingerprints collected from 15different subjects. We followed the dataset split strategyproposed in [30], randomly choosing 395 images as galleryand 520 as probes, making sure that each class contained

at least one gallery sample. 131 images were left out sinceSankaran et al. used them for training. We performed 10-fold cross-validation in order to ensure the random splittingdoes not influence the reported performance.

Table 1, shows Rank-1 and Rank-10 accuracy for bothrecognition methods with and without our enhancement.Our approach significantly improves the matching accuracy.Comparing to [30], we outperform all the fingerprint match-ing methods they have evaluated. It is worth emphasizingthat differently from Sankaran et al. we do not need to trainon the subset of the data as they do.

The CMC curves for this experiment are shown in Fig-ure 3. Our model boosts the performance of ABR11 + MCCmuch more than that of NBIS. We attribute this to the factthat the energy we minimize while training our model per-forms very similar operations as the ridge binarization partof the ABR11 feature extraction algorithm.

Latent-to-Sensor matching. Our second set of experi-ments tackled an even more challenging task. We used theMOLF dataset containing all ten fingerprints of 100 dif-ferent subjects. The samples in this dataset were of verydifferent quality, including some very poor samples whereno ridge structure was visible at all. Fingerprints of eachparticipant were captured with several commercial finger-print scanners (Lumidigm, Secugen and Crossmatch). Inaddition, each participant provided a set of latent finger-prints. It is therefore possible to match latent fingerprintsto those acquired by a sensor. Following the testing pro-tocol by Sankaran et al. [31], we considered the first andsecond instance fingerprints for each user from a sensorscanned database as the gallery. The whole latent finger-print database consisting of 4400 samples was consideredas probe set.

For this experiment, we refer to the case from [31],where minutiae were extracted automatically and after-wards matched using one of the standard algorithms.Sankaran et al. evaluated the performance of publicly avail-able NBIS and commercial VeriFinger [26] fingerprintmatching methods, reporting very poor performance forboth. Here, we used the ABR11 + MCC method for match-ing. Table 2 shows that using our enhancement algorithm,

5 10 15 20 2520

40

60

80

100

Rank-N

Acc

urac

y(%

)

MINDTCT+BOZABR11+MCCOUR+MINDTCT+BOZOUR+ABR11+MCC

Figure 3. Latent-to-Latent matching on IIIT-Delhi latent database.CMC curve showing the performance improvement of our en-hancement approach (solid lines) against matching raw latent fin-gerprints (dashed lines).

Method AccuracyEnhancement Extract + Match Rank-1 Rank-10

Raw ABR11 + MCC 35.58% 58.27%Raw NBIS 57.69% 71.92%Our ABR11 + MCC 57.12% 79.04%Our NBIS 62.69% 78.85%

Table 1. Latent-to-Latent matching on IIIT-Delhi latent database.Matching was performed on images obtained with the proposedfingerprint enhancement (Our) and with no enhancement (Raw).

10 20 30 40 500

10

20

30

Rank-N

Acc

urac

y(%

)

MINDTCT+MCCABR11+MCCOUR+MINDTCT+MCCOUR+ABR11+MCC

Figure 4. Latent-to-Sensor matching on IIIT-Delhi MOLFdatabase. CMC curves showing comparison of our enhancementapproach (solid lines) against matching raw latent fingerprints(dashed lines) to the Lumidigm database.

ABR11 + MCC performs very poorly, consistently withthe finding of [31]. Performing the enhancement withour model significantly improves the performance. CMCcurves for this experiment are shown Figure 4.

Method AccuracyEnhancement Extract + Match Rank-25 Rank-50

Raw MINDTCT + MCC 8.02% 12.59%Raw ABR11 + MCC 3.07% 5.45%Our MINDTCT + MCC 12.55% 18.36%Our ABR11 + MCC 16.14% 22.36%[31] NBIS N/A 6.06%[31] VeriFinger N/A 6.80%

Table 2. Latent-to-Sensor matching on IIIT-Delhi MOLFdatabase. Shown as the performance of matching Lumidigm im-ages to latent fingerprints with enhancement (Our) and with noenhancement (Raw). Two more methods listed in [31] are com-pared.

10 20 30 40 500

10

20

30

Rank-N

Acc

urac

y(%

)

Raw-CrossmatchRaw-SecugenRaw-LumidigmOur-CrossmatchOur-SecugenOur-Lumidigm

Figure 5. Cross-dataset Latent-to-Sensor matching on IIIT-DelhiMOLF database. CMC curve showing the performance improve-ment of our enhancement approach (solid lines) against matchingraw latent fingerprints (dashed lines) to the Lumidigm, Secugen,and Crossmatch scanner datasets.

Method AccuracyEnhancement Gallery dataset Rank-25 Rank-50

Raw Lumidigm 3.07% 5.45%Raw Secugen 2.59% 5.18%Raw Crossmatch 2.68% 5.32%Our Lumidigm 16.14% 22.36%Our Secugen 13.27% 19.50%Our Crossmatch 12.66% 19.07%

Table 3. Cross-dataset Latent-to-Sensor matching on IIIT-DelhiMOLF database. Performance of matching Lumidigm, Secugen,and Crossmatch sensor images to latent fingerprints enhanced byour model (Our) and with no enhancement (Raw). In all cases,combination ABR11 + MCC was used for feature extraction andmatching.

Cross-dataset Latent-to-Sensor matching. To supportthe fact that our method works not only matching latent fin-gerprints to Lumidigm (MOLF DB1) sensor scanned finger-prints, we provide a comparison of matching latent finger-

prints to Secugen (MOLF DB2) and Crossmatch (MOLFDB3) sensor samples as well using the combination ofABR11 feature extractor and MCC matching method. Assamples from various sensors are of different quality, theperformance can differ slightly. However, Table 3 indicatesthat independently of the source sensor, our enhancementalgorithm provides significant performance improvement.The CMC curves for this experiment are shown in Figure5.

Fingerprint quality. In order to qualitatively demonstratehow well our model reconstructs the latent fingerprints, weperform quality assessment using the NFIQ utility (part ofNBIS [27]), which assigns a fingerprint image a numericalscore from 1 (best quality) to 5 (poorest quality). The dis-tribution of score values shown in Figure 6 clearly indicatesthat our method improves the fingerprint quality.

Moreover, we show several successful and several un-successful latent fingerprint reconstructions for both real(Figure 7) and synthetic (Figure 8) data. It is clearly visiblethat our model enhances the ridge information well whenit is present and has problems in places where the originalimages does not contain discernible ridges. This suggeststhat very poor samples might not contain any meaningfulinformation for the reconstruction process.

1 2 3 4 5

0

1,000

2,000

3,000

4,000

Quality Score

Num

bero

fim

ages

RawAdapt. Hist. Eq.Our

Figure 6. Image quality assessment using NFIQ utility. We com-pare the raw data with data improved using adaptive histogramequalization and data improved using our enhancement. The qual-ity assessment score ranges from 1-5, where the lower the better.

6. Conclusion

Convolutional autoencoders are amongst popular meth-ods extensively used in image processing tasks for imagedenoising and inpainting. Inspired by the previous applica-tions, we have explored the possibility of using the convolu-tional autoencoders to reconstruct latent or damaged finger-prints. With regard to principles of some of the fingerprint

(a) Successful reconstructions

(b) Fail cases

Figure 7. Examples of the fingerprint reconstructions on real la-tent fingerprints. Each pair is (input, reconstruction) (a) success-fully reconstructed samples (b) cases where the model fails to re-construct the fingerprint well.

(a) Successful reconstructions (b) Fail cases

Figure 8. Examples of the fingerprint reconstructions on synthetic data. Each triplet is from left to right (input, reconstruction, target) (a)successfully reconstructed samples (b) cases where the model fails to reconstruct the fingerprint well.

feature extraction and matching algorithms, we have care-fully designed an objective function that should both wellreflect on the important fingerprint properties and be effi-cient to optimize as it is based on gradient analysis, whichhas been already implemented on the GPUs. Our method isbased on learning, however, in comparison to some of theprevious research, we do not need any real training data andwe effectively train our model on well designed syntheticdataset, which gives us the advantage of no training set sizelimitation.

We obtain state-of-the-art results on several challeng-ing tasks such as latent-to-latent fingerprint matching andlatent-to-sensor database fingerprint matching on IIIT-Dstandard datasets, outperforming the existing results usingon the same data by a margin. Especially evaluation on verychallenging IIIT-D MOLF dataset compensates for the factthat we cannot evaluate on NIST-SD27 as it is discontinuedand no longer provided by NIST.

On the other hand, we observe that the reconstruction isnot always successful, showing some of the failure casesthat are prone to generate false minutiae. We would like tofurther examine this issue as future work.

As we do not aim to directly extract minutiae or per-

form matching, but rather reconstruct the correct ridge pat-tern of the poor-quality fingerprint, our method has broadapplications ranging from latent fingerprint enhancement topossibly reconstruction of fingerprints affected by diseases,which is another of our desired future directions.

7. Acknowledgements

The authors are supported in part by ERC Starting GrantNo. 307047 (COMET), ERC Consolidator Grant No.724228 (LEMAN) and Nvidia equipment grant.

References[1] J. Abraham, P. Kwan, and J. Gao. Fingerprint Matching us-

ing A Hybrid Shape and Orientation Descriptor. State of theart in Biometrics. InTech, 2011.

[2] A. H. Ansari. Generation and storage of large synthetic fin-gerprint database. Master’s thesis, IISc, 7 2011. M.E. Thesis.

[3] D. R. . Ashbaugh. Handbook of Fingerprint Recognition.CRC Press, 1999.

[4] J. S. Bartunek, M. Nilsson, J. Nordberg, and I. Claesson.Adaptive fingerprint binarization by frequency domain anal-ysis. In Proc. ACSSC, pages 598–602, Oct 2006.

[5] K. Cao, E. Liu, and A. K. Jain. Segmentation and enhance-ment of latent fingerprints: A coarse to fine ridgestructuredictionary. PAMI, 36(9):1847–1859, Sept 2014.

[6] R. Cappelli, M. Ferrara, and D. Maltoni. Minutia cylinder-code: A new representation and matching technique for fin-gerprint recognition. PAMI, 32(12), 2010.

[7] R. Cappelli, M. Ferrara, and D. Maltoni. Fingerprint in-dexing based on minutia cylinder-code. PAMI, 33(5):1051–1057, 2011.

[8] R. Cappelli, D. Maio, and D. Maltoni. Synthetic fingerprint-image generation. pages 475–478, 2000.

[9] A. Dosovitskiy, J. T. Springenberg, and T. Brox. Learning togenerate chairs with convolutional neural networks. CoRR,2014.

[10] J. Feng, J. Zhou, and A. K. Jain. Orientation field estimationfor latent fingerprint enhancement. PAMI, 35(4):925–940,April 2013.

[11] M. Ferrara, D. Maltoni, and R. Cappelli. Noninvertibleminutia cylinder-code representation. IEEE Trans. on Infor-mation Forensics and Security, 7(6):1727–1737, 2012.

[12] M. Ferrara, D. Maltoni, and R. Cappelli. A two-factorprotection scheme for mcc fingerprint templates. In Proc.BIOSIG, pages 1–8, 2014.

[13] S. Hong, H. Noh, and B. Han. Decoupled deep neural net-work for semi-supervised semantic segmentation. CoRR,2015.

[14] M. K. Hu. Visual Pattern Recognition by Moment Invariants.IRE Trans. on Information Theory, IT-8:179–187, Feb. 1962.

[15] S. Ioffe and C. Szegedy. Batch normalization: Acceleratingdeep network training by reducing internal covariate shift.CoRR, 2015.

[16] A. K. Jain and J. Feng. Latent fingerprint matching. PAMI,33(1):88–100, 2011.

[17] M. S. Khalil. Deducting fingerprint singular points usingorientation field reliability. In Proc. RVSP, pages 284–286,Nov 2011.

[18] D. P. Kingma and J. Ba. Adam: A method for stochasticoptimization. CoRR, 2014.

[19] P. Komarinski. Automated Fingerprint Identification Systems(AFIS). Academic Press, 2004.

[20] H. C. Lee, R. Ramotowski, and R. E. Gaensslen. Advancesin Fingerprint Technology. CRC Press, 2001.

[21] A. L. Maas, A. Y. Hannun, and A. Y. Ng. Rectifier nonlin-earities improve neural network acoustic models. In Proc.ICML, volume 30, 2013.

[22] D. Maltoni, D. Maio, A. K. Jain, and S. Prabhakar. Hand-book of Fingerprint Recognition. Springer Publishing Com-pany, Incorporated, 2nd edition, 2009.

[23] X. Mao, C. Shen, and Y. Yang. Image restoration using con-volutional auto-encoders with symmetric skip connections.CoRR, 2016.

[24] J. Masci, U. Meier, D. Ciresan, and J. Schmidhuber. StackedConvolutional Auto-Encoders for Hierarchical Feature Ex-traction, pages 52–59. Springer Berlin Heidelberg, Berlin,Heidelberg, 2011.

[25] V. Nair and G. E. Hinton. Rectified linear units improve re-stricted boltzmann machines. In J. Frnkranz and T. Joachims,editors, Proc. ICML, pages 807–814. Omnipress, 2010.

[26] NeuroTechnology. Verifinger.www.neurotechnology.com/verifinger.html.

[27] NIST. Nbis (nist biometric image software).http://www.nist.gov/itl/iad/ig/nbis.cfm.

[28] H. Noh, S. Hong, and B. Han. Learning deconvolution net-work for semantic segmentation. CoRR, 2015.

[29] A. Radford, L. Metz, and S. Chintala. Unsupervised repre-sentation learning with deep convolutional generative adver-sarial networks. CoRR, 2015.

[30] A. Sankaran, T. I. Dhamecha, M. Vatsa, and R. Singh. Onmatching latent to latent fingerprints. In Proc. IJCB, pages1–6, Oct 2011.

[31] A. Sankaran, M. Vatsa, and R. Singh. Multisensor opticaland latent fingerprint database. IEEE Access, 3:653–665,2015.

[32] P. Schuch, S. Schulz, and C. Busch. De-convolutional auto-encoder for enhancement of fingerprint samples. In Proc.IPTA, pages 1–7, Dec 2016.

[33] S. Yoon, J. Feng, and A. K. Jain. On latent fingerprint en-hancement. 2010.

[34] S. Yoon, J. Feng, and A. K. Jain. Latent fingerprint enhance-ment via robust orientation field estimation. pages 1–8, 2011.

[35] J. J. Zhao, M. Mathieu, R. Goroshin, and Y. LeCun. Stackedwhat-where auto-encoders. CoRR, 2015.