Sei sulla pagina 1di 41

FINDINGS ON OFFLINE

HANDWRITTEN SIGNATURE
VERIFICATION
Paper-1: Offline signature verification with Convolutional Neural
Network
~Stanford University

 Method used: Convolutional Neural Network (CNN)


 Model based on: VGG16 architecture
 Dataset used: ICDAR 2011 SigComp datset
 Training method: Transfer learning
 Accuracy: 97% for Dutch signatures
95% for Chinese signatures
Paper-2: Learning features for Offline Handwritten Signature Verification using Deep
Convolutional Neural Network (16 May 2017)
~University of Quebec, Montreal, Quebec, Canada
Federal University of Parana (UFPR), Curitiba, PR, Brazil

 Method used: Convolutinal Neural Network (CNN)


 Datasets used: GPDS, MCYT, CEDAR, Brazilian PUC-
PR
 Accuracy: On GPDS-160 – 1.72% Equal Error Rate,
compared to 6.97% in the literature
 Training method: Feature learning
Paper-3: Offline Handwritten Signature identifier using Adaptive Window Positioning
Techniques
~Universiti Tecknologi Malaysia (UTM) Skudai 81310, Johor, Malaysia

 Method used: Adaptive window positioning technique for feature extraction


- Division of signature images into 13x13 windows and created some new cluster
patterns for each window when classified into groups of similar attributes
 Dataset used: GPDS dataset containing 4870
signatures samples from 90 different writers
 Results:
Paper-4: A writer-independent approach for offline signature
verification using deep convolutional neural networks features
(26 July 2018)

 Method used: Features extracted using a deep


Convolutional Neural Network (CNN), based on the
dichotomy transformation combined with a writer-
dependent (WD) Support Vector Machine (SVM)
classifier
 Dataset used: Brazilian and GPDS datasets
 Results:
Paper-5: Offline signature verification by combining Graph
Edit Distance and Triplet Networks (17 Oct 2018)

 Method used: Combining structural and statistical models.


The structural model is based on approximate Graph Edit
Distance. The statistical model is on metric learning with
deep triplet networks.
 Note: “We can see several lines of future research. For the
structural method, more graph-based representations and
cost functions may be explored in the context of graph edit
distance. For the statistical method, synthetic data
augmentation may lead to a more accurate vector space
embedding. Finally, we believe that there is a great
potential in combining even more structural and statistical
classifiers into one large multiple classifier system. Such a
system is expected to further improve the robustness of
biometric authentication.”
 Dataset used: MCYT and GPDS datasets
 Training method: Deep CNN, using the ResNet18
architecture. Three different models are trained
using the DeepDIVA framework.
 Results:
Paper-6: Writer Independent Offline Signature
recognition using Ensemble Learning (19 Jan 2019)

 Method used: Ensemble model with Deep Learning.


Two CNNs for feature extraction, after that
Regularized Gradient Boosting Tree (RGBT) for
classification and Stacking as an Ensemble
algorithm to generate final prediction vector
 Dataset used:
 Results:
Paper-7: Graph based Offline signature
verification (25 June 2019)
 Method used: Two graph-based approaches:
Keypoint graphs with approximated graph edit
diatance and inkball models
 Dataset used:
 Results:
 Note: “The performance of the proposed systems when
encountering random forgeries is not as strong as the state
of the art. As proposed by [29], a combination with a neural
network should further improve the performance of the
structural approach on random forgeries. The results also
show that the dierence between the user-specic and the
global EER is quite large. This indicates that the proposed
approaches could benet from improved user adaptation in
the future. Overall, the results are quite remarkable,
especially when considering that the proposed systems are
applied on four dierent test sets without any further
adaptation while being trained on synthetic signatures only.”
Paper-8: SigNet: Convolutional Siamese Network for Writer
Independent Offline Signature Verification
(30 Sept 2017)

 Method used: Convolutional Siamese Network, using


a CNN architecture
 Datasets used: CEDAR, GPDS300, GPDS Synthetic,
BHSig260
 Results:
Paper-9: Handwritten Signature Forgery
Detection using Convolutional Neural Networks

 Method used: CNNs, particularly ACNN (or


ConvNet), where an image goes through 3
convolution and max pooling layers which are in an
alternating fashion
 Dataset used:
 Results:
Paper-10: Handwritten Signature
Verification using Neural Network
 Method used: Image pre-processing, geometric
feature extraction, neural network training with
extracted features and verification
 Dataset used: 2106 signatures containing 936
genuine and 1170 forgeries
 Results:
Paper-11: Signature Verification using
a Convolutional Neural Network
 Method used: Convolutional Neural Network (CNN)
 Dataset used: SIGCOMP 2011 dataset
 Results:
Paper-12: An Efficient Signature Verification Method based on
an Interval Symbolic Representation and a Fuzzy Similarity
Measure

 Method used: Based on an interval symbolic representation and a fuzzy


similarity measure is proposed. In the feature extraction step, a set of Local
Binary Pattern (LBP) based features is computed from both the signature
image and its under-sampled bitmap.
 Dataset used: Benchmark off-line English signature dataset
(GPDS-300) and a large dataset (BHSig260) composed of
Bangla and Hindi off-line signatures .
 Results:
Paper-13: Wavelet-based Off-line
Signature Verification (1999)
 Method used: “In this paper, a wavelet–based off–line signature verification system
is proposed. The system starts with a closed–contour tracing algorithm. The
curvature data of the traced closed contours are decomposed into multiresolutional
signals using wavelet transforms. Then the zero–crossings corresponding to the
curvature data are extracted as features for matching. Moreover, a statistical
measurement is devised to decide systematically which closed contours and their
associated frequency data of a writer are most stable and discriminating. Based on
these data, the optimal threshold value which controls the accuracy of the feature
extraction process is calculated.”
 Dataset used: “We collected genuine signatures from 25 English writers and 25
Chinese writers, each of whom wrote their signature 20 times. Among the 20
genuine signatures of each writer, 10 were used as templates, and the other 10
were used for testing. We also collected 5 different quality types of forged
signatures from 40 senior undergraduate students during one semester, where 10
forgery signatures were collected per each forgery signature type from each
writer. Therefore, there were 3,500 signatures in total in the database.”
 Results: “Experimental results show that the average success rates for English
signatures and Chinese signatures are 91.71% and 93%, respectively.”
 “For the English signature, the proposed system mis–
identified about 21.20% of forged signatures as
genuine ones. As for the Chinese signature, the mis–
identification rate was about 13.50%.”

Potrebbero piacerti anche