Deep Feature-Based Writer-Dependent Classifiers for Offline Signature Verification

In this work, we proposed a novel scheme based on deep architectures for offline signature verification. The proposed method introduces the notion of writer-dependent deep architectures for offline signature verification. Compared to the current signature verification techniques that use the same architecture for all writers, the proposed model based on applying deep architecture which may vary from a writer to writer. In this work, writer-dependency has been exploited at two stages: In the first stage, writer-dependent deep architectures are selected for each writer. In the second stage, writer-dependent deep architectures are used as feature extractors, and then the dimensionality of a feature vector is reduced through the application of linear dimensionality reduction technique. Finally, writerdependent classifiers are fixed for each writer. At the verification stage, to establish the authenticity of the test signature, features are extracted from the writer-dependent architecture and fed into the writer-dependent classifier of the claimed writer. Extensive experiments are carried out on two benchmark offline signature datasets: CEDAR and MCYT, to validate the performance of the proposed model. The obtained results clearly indicate the efficacy of the proposed methodology.