Igor Janos（Innovatrics）：利用合成數據突破潛在指紋識別的極限.pdf

《Igor Janos（Innovatrics）：利用合成數據突破潛在指紋識別的極限.pdf》由會員分享，可在線閱讀，更多相關《Igor Janos（Innovatrics）：利用合成數據突破潛在指紋識別的極限.pdf（29頁珍藏版）》請在三個皮匠報告上搜索。

1、IFPC 2025 ConferenceIgor JnoHead of R&DPushing the limits of latent fingerprint identification with synthetic data 2MotivationImages of latent fingerprints are scarceOften they are confidentialpart of an investigation(ongoing or closed)Matching pairs(latent-plain)are difficult to get3Known Fingerpri

2、ntsAcquired using ink or digital scannersControlled environmentConstant DPILess diverse4Unknown FingerprintsDiverse backgroundsPartial informationAcquired using different techniques5Annotating latent fingerprints is difficultHaving forensic experts do the annotation is expensiveLaymen will not alway

3、s be consistent6Our Approach7Our ApproachTrain on purely synthetic latent imagesWe acknowledge lot of research on synthetic fingerprint generationCycle-GAN-based approachesDiffusion-based approachesProcedural approaches8TrainingPredict the same features from plain and latent imagesPlain Feature Extr

4、actorLatent Feature ExtractorLoss between featuresGround TruthPrediction9Training DataDomain A(Plains)Domain B(Latents)Produce big bag of image pairsAugmentationGANsDiffusionWe have no real matching pairsUnpaired image translation10Training DataReal PlainsReal Latents(SD302)Latents contain a lot les

5、s information than the plains11Submission 112Reconstructed PlainsOur Submission 1Masked by random polygonAdded random noiseReal PlainsP LSynthetic LatentsLatent DomainDiscriminatorL PConsistency Loss13Reconstructed LatentsOur Submission 1Real Latents still containa lot less informationReal LatentsL

6、PSynthetic PlainsPlain DomainDiscriminatorP LConsistency Loss14Reconstructed LatentsOur Submission 1Plain Discriminator is forcingthe(LP)model to make up informationReal LatentsL PSynthetic PlainsPlain DomainDiscriminatorP LConsistency Loss15Reconstructed LatentsOur Submission 1In reality,this is no

7、t a bijective projection.Leads to overfitting.Real LatentsL PSynthetic PlainsPlain DomainDiscriminatorP LConsistency Loss1617Our Submission 1innovatrics-0005Not great,not terrible resultsNutrition Report,Detection Error Tradeoff(DET)Nutrition Report,Cumulative Match Characteristic(CMC)18Easy PlainDi

8、fficult Latent19Submission 220Our Submission 2Preserve discriminative informationBe creative where little information is stored Avoid hard-coded clipping mask21DecoderOur Submission 2EncoderTrain on inpainting22DecoderOur Submission 2EncoderEncoder understands the shape,high-level features,ridges,or

9、ientation23MinutiaeRegressorOur Submission 2FrozenEncoderNow train a minutiae regressor2425Nothing important is hereMinutiae are preserved26Our Submission 2innovatrics-0009Kind-of promisingNutrition Report,Detection Error Tradeoff(DET)Nutrition Report,Cumulative Match Characteristic(CMC)27Current St

10、ate28Current StateImproved techniques of data synthesisStill trained on purely synthetic datainnovatrics_0009innovatrics_000DNutrition Report,Cumulative Match Characteristic(CMC)2020 Innovatrics.All Rights Reserved Slovakia(HQ)+421 2 2071 4056Brazil+55 11 4210-5185Singapore+65 3158 7379Taiwan(R.O.C.)+886 2 7741 4036USA+1 404 984-2024Thank you