Estimation of Fingerprint Orientation Fields and Deep Expectation Method

i-VisionGroup Deep Expectation for Estimation of Fingerprint Orientation Fields Patrick Schuch Simon-Daniel Schulz Christoph Busch IJCB17 2017-11-17 i-VisionGroup

Intro DEX: Deep Expectation 2015 ICCV workshop 2016 IJCV Deep Expectation of Real and Apparent Age from a Single Image Without Facial Landmarks classification ConvNets can be superior to regression ConvNets even on such regression problems 2 i-VisionGroup

Intro Fingerprint Orientation Estimation (FOE): Central processes in feature extraction for fingerprints Challenging on low quality 3 i-VisionGroup

Intro Benchmark: FVC-ongoing Deep learning has provided promising results 4 i-VisionGroup

Intro Five questions regarding ConvnetOF whether the proposed use of Deep Expectation can outperform Regression or Naive Classification how generalization can be achieved despite this imbalance whether there is a better layout of the network for FOE an appropriate dimensioning for a ConvNet for FOE How the foreground masking provided by FVC-ongoing improves performance 5 i-VisionGroup

Proposed method Deep Expectation: In order to achieve good regression performance using a classification approach, the regression nature can be encoded by establishing a relation between the classes. The idea is to avoid a winner-takes-all policy by using a weighting strategy 6 i-VisionGroup

Proposed method Double angle for phase shift 7 i-VisionGroup

Experiments & Results Layout of ConvNetOF 8 i-VisionGroup

9 i-VisionGroup

Q1:Prediction Mode Regression or classification or DEX? Regression: Original layout of ConvNetOF Naive Classification: Layout like ConvNetOF but the final output layer has 256 neuron - one for each orientation class. The training costs were based on cross-entropy of the 256 output neurons Implicit Deep Expectation: We took the trained Naive Classification model and estimated the fingerprint orientation field by Deep Expectation when testing. Explicit Deep Expectation: Same layout as Naive Classification but the training cost function depended directly on the orientation estimation carried out in a Deep Expectation manner. 10 i-VisionGroup

11 i-VisionGroup

Q2:Regularization Limited training data Additional data: worse performance (discrepancies in marking the ground truth) Some common methods for regularization Weight Decay Drop out Affine Data-Augmentation Batch Normalization & Batch Size Gaussian Noise Gamma Correction 12 i-VisionGroup

13 i-VisionGroup

Q3:Model layout Inception Modules: from GoogLeNet 3x3 Cascades: replace larger kernels Dimension Reduction: add 1x1 convlayers AlexNet: dropped the last pooling layers 14 i-VisionGroup

3x3 cascades: too bad performance 15 i-VisionGroup

Q4:Model size Especially the GP layers and the FC layers seemed to be oversized Several attempts: 50% of all Neurons 50% Fully Connected Slightly Smaller GPs: 32 & 11x11 instead of 49 & 13x13 2 GP Blocks 4 GP Blocks 16 i-VisionGroup

17 i-VisionGroup

Q5:Mask 18 i-VisionGroup

Final: DEX-OF 19 i-VisionGroup

Thanks for listening Q&A 20 i-VisionGroup