Open AccessDeepFuseOSV: online signature verification using hybrid feature fusion and depthwise separable convolution neural network architecture
Author(s) -
Sekhar Vorugunti Chandra,
Pulabaigari Viswanath,
Mukherjee Prerana,
Sharma Abhishek
Publication year - 2020
Publication title -
iet biometrics
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.434
H-Index - 28
eISSN - 2047-4946
pISSN - 2047-4938
DOI - 10.1049/iet-bmt.2020.0032
Subject(s) - computer science , feature (linguistics) , artificial intelligence , convolution (computer science) , convolutional neural network , signature (topology) , encoder , word error rate , pattern recognition (psychology) , deep learning , artificial neural network , philosophy , linguistics , geometry , mathematics , operating system
Online signature verification (OSV) is a widely utilised technique in the medical, e‐commerce and m‐commerce applications to lawfully bind the user. These high‐speed systems demand faster writer verification with a limited amount of information along with restrictions on training and storage cost. This study makes two major contributions: (i) A competent feature fusion technique in which traditional statistical‐based features are fused with deep representations from a convolutional auto‐encoder; and (ii) a hybrid architecture combining depth‐wise separable convolution neural network (DWSCNN) and long short term memory (LSTM) network delivering state‐of‐the‐art performance for OSV is proposed. DWSCNN is utilised for extracting deep feature representations and LSTM is competent in learning long term dependencies of stroke points of a signature. This hybrid combination accomplishes better classification accuracy (lower error rates) even with one‐shot learning, i.e. achieving higher classification accuracies with only one training signature sample per user. The authors have extensively evaluated their model using three widely used datasets MCYT‐100, SVC and SUSIG. These exhaustive experimental studies confirm that the DeepFuseOSV framework results in the state‐of‐the‐art outcome by achieving an equal error rate (EER) of 13.26, 2.58, 0.07% in Skilled 1, Skilled 10 and Random 10 categories of MCYT‐100, respectively, 7.71% in Skilled 1 category of SVC, 1.70% in Random 1 category of SUSIG.