Writer-independent Feature Learning for Offline Signature Verification using Deep Convolutional Neural Networks

I. INTRODUCTION

• Offline verification uses a scanned signature image rather than dynamic information captured during the writing process.Signatures are important biometric traits because they are widely used for identity verification in legal, financial, and administrative settings.Source: §I · p. 1
• Systems still struggle to distinguish genuine signatures from skilled forgeries, which can closely resemble genuine samples despite substantial variation among a writer’s genuine signatures.The challenge is especially pronounced when only a few genuine samples are available for training.Source: §I · p. 1, Figure 1 · p. 2
• The paper proposes learning signature representations from data with supervised CNNs instead of relying on hand-crafted features that may not reflect how signatures are created.CNNs scale better than fully connected models for larger inputs, with fewer trainable parameters, helping retain details that distinguish skilled forgeries.Source: §I · pp. 1–2
• A writer-independent CNN learns features from users outside the enrolled system, after which writer-dependent classifiers are trained for individual users.Adding a user then requires training only that user’s writer-dependent classifier.Source: §I · p. 2
• The study tests the method on GPDS-960 and Brazilian PUC-PR and reports state-of-the-art performance and investigations of cross-dataset transfer and training-sample counts.GPDS-960 is described as the largest publicly available offline signature-verification corpus; PUC-PR is a smaller dataset used in prior studies.Source: §I · p. 2

II. RELATED WORK

• Murshed et al., used single-hidden-layer autoencoders for dimensionality reduction, while Ribeiro et al. evaluated RBM representations on only 10 users without reporting discrimination results.The text contrasts these approaches with deeper, often over-complete autoencoders intended to disentangle input variation.Source: §II · p. 2
• Khalajzadeh applied CNNs to Persian signature verification but did not consider skilled forgeries.Source: §II · p. 2
• Sun et al. transferred CNN features between face datasets, while Eskander et al. used writer-independent data for feature selection rather than feature learning.The present work evaluates both random and skilled forgeries, which the cited face-verification strategy did not address as its main challenge.Source: §II · p. 2

III. PROPOSED METHOD

• The dataset is partitioned into development set D for learning representations and exploitation set E for enrolled users’ classifier training and evaluation.The development set is treated as separate from the enrolled users.Source: §III · p. 2
• A CNN trained on D learns a function φ that maps signature images into an m-dimensional feature space.The method expects these features to help separate genuine signatures and forgeries from other users.Source: §III · pp. 2–3, Figure 2 · p. 3
• For each user in E, the CNN extracts features from genuine signatures and random forgeries to train a binary classifier that decides whether a new signature is genuine or forged.At inference, a new image is mapped by φ and its feature vector is passed to the user-specific classifier.Source: §III · pp. 2–3

A. Pre-processing

• Images are resized to a fixed neural-network input size, either directly with aspect-ratio-preserving cropping or after padding and centering them on an 840 x 1360 canvas.The padded approach aligns each signature’s center of mass to the canvas center before resizing.Source: A. Pre-processing · p. 3
• Direct resizing loses less fine-grained information, especially for small signatures, but makes their pen strokes thicker relative to strokes in larger signatures.Source: A. Pre-processing · p. 3
• The pipeline removes background pixels with OTSU thresholding and inverts images so the white background has intensity zero.Pixels above the threshold are set to white before inversion.Source: A. Pre-processing · p. 3
• Input pixels are divided by the standard deviation of pixel intensities across D, without mean-centering, so background pixels remain zero-valued.Source: A. Pre-processing · pp. 3–4

B. Writer-Independent feature learning

• The CNN learns features by classifying each development-set user as a separate class, using cross-entropy as a surrogate objective for the target verification task.The target users’ skilled forgeries are unavailable during enrollment, motivating this separate-user classification task.Source: B. Writer-Independent feature learning · p. 4
• The architecture follows Krizhevsky et al. but uses two fully connected layers after the convolutions instead of three, after initial tests suggested excess capacity.The authors identify the fully connected layers as containing most of the network’s weights.Source: B. Writer-Independent feature learning · p. 4
• Training uses Nesterov momentum for 60 epochs, batches of 100, a learning rate reduced tenfold after epochs 20 and 40, and L2 weight decay of 0.0005.The weights are initialized according to Glorot and Bengio, and the biases are initialized to zero.Source: B. Writer-Independent feature learning · p. 4

C. Writer-dependent classification

• For each signature, the CNN supplies a 4096-dimensional representation from the layer immediately before softmax, which is then used for writer-dependent classification.The representation is denoted φ(X).Source: C. Writer-dependent classification · p. 4
• Writer-dependent training excludes skilled forgeries and uses genuine signatures plus random forgeries drawn from other users as negative examples.The protocol is intended to simulate an application setting where only enrolled users’ genuine signatures are available.Source: C. Writer-dependent classification · p. 4
• For each user, a two-class classifier is trained on genuine signatures from that user and genuine signatures from development-set users, then tested on held-out genuine signatures and that user’s skilled forgeries.The CNN first extracts features for each signature in the user-specific training and testing sets.Source: C. Writer-dependent classification · p. 4
• The authors test linear and RBF-kernel SVMs; the RBF parameters C and γ are selected by grid search on development-set users.For a new signature, the CNN representation is passed to the user-specific classifier for a final decision.Source: C. Writer-dependent classification · pp. 4–5
• The GPDS-960 dataset is divided into development set D and exploitation set E for feature learning and enrolled-user classification, respectively.The supplied figure caption identifies this D/E separation; it does not specify the split sizes.Source: Figure 3 · p. 5

IV. EXPERIMENTAL PROTOCOL

• GPDS-960 contains 881 users with 24 genuine signatures and 30 forgeries each, while Brazilian PUC-PR contains 168 users and forgeries for the first 60.GPDS signatures were captured in a single session; the GPDS exploitation set uses the first 160 or 300 users, with the remaining users in development, while Brazil uses the first 60 for exploitation and the remaining 108 for development.Source: §IV · p. 5
• Images are preprocessed to 155 x 220 pixels, and the development set trains a CNN to classify its users.The size was chosen to preserve pen-stroke detail while remaining small enough for GPU training.Source: §IV · p. 5
• The authors also extract Brazilian-dataset features with a CNN trained on GPDS to assess cross-dataset generalization and the use of the larger dataset for the smaller one.GPDS was selected as the larger feature-learning dataset.Source: §IV · p. 5
• The writer-dependent protocols differ by dataset: GPDS uses up to 14 user signatures for training and tests on 10 held-out genuine signatures plus 30 skilled forgeries; Brazil uses up to 30 for training and tests on 10 genuine, 10 random, 10 simple, and 10 skilled forgeries.Table III is identified as the training and testing set-up.Source: §IV · p. 5, Table III · p. 6
• The number of enrolled-user signatures is varied, and SVM hyperparameters are tuned on 10 development-set users; inputs are standardized and genuine examples duplicated to balance training classes.The authors report that standardization slightly improved performance and significantly reduced SVM training time.Source: §IV · p. 5
• GPDS results use FRR, skilled-forgery FAR, EER, and mean AUC; Brazilian results include FRR, FAR for random, simple, and skilled forgeries, and AER.For comparison with GPDS, the Brazilian evaluation also reports AERgenuine + skilled, EERgenuine + skilled, and Mean AUCgenuine + skilled.Source: §IV · pp. 5–6

• Normalization improved GPDS-160 EER from 14.64% to 10.70%, and the RBF SVM performed best.The selected settings were γ = 2^-12 and C = 1, the best configuration for 8/10 development users.Source: §V · p. 6
• CNNs trained on the Brazilian and GPDS development sets yielded similar PUC-PR results, despite development sets of 108 and 721 users, respectively.Source: §V · p. 6
• 15.05% EER with 4 GPDS signatures and 9.83% with 5 Brazilian signatures contrasted with around 17% EER when only one signature was available.Source: §V · p. 6, Figure 4 · p. 6, Figure 5 · p. 7
• On GPDS, performance is state-of-the-art against single-feature-extractor systems but below multi-extractor systems; Brazilian PUC-PR results are state-of-the-art.Source: §V · pp. 6–7, Table VII · p. 7, Table VI · p. 7
• A global threshold appears insufficient: GPDS had large FRR with low FAR, whereas the Brazilian dataset showed the opposite pattern.The authors suggest considering user-specific thresholds.Source: §V · p. 7
• Writer-dependent training combines genuine signatures with random forgeries under an explicitly weak proxy assumption, and the training sets are highly imbalanced.The GPDS-160 setup had 1–30 positive samples and up to 10,000 random forgeries.Source: §V · p. 7

VI. CONCLUSION

• The method performs close to state of the art on GPDS and surpasses it on Brazilian PUC-PR, with features learned on GPDS generalizing to PUC-PR.The authors also report effectiveness with 4–5 samples per user.Source: §VI · p. 8
• The method can be effective with few samples per user, including 4–5 samples.Source: §VI · p. 8
• Despite low Equal Error Rates, false rejection and false acceptance rates are highly imbalanced and unstable across users and datasets; defining user-specific thresholds remains future work.Source: §VI · p. 8