Discriminative DCT-MLP based approach for off-line signature verification

In this paper, we propose to explore the possibility of integrating the appearance based paradigm with frequency domain for off-line signature verification. The proposed approach has four major phases : Preprocessing, Feature extraction, Feature reduction and Classification. In the feature extraction phase, Discrete Cosine Transform (DCT) is employed on the signature image to obtain the upper-left corner block of size m × n as a representative feature vector. These features are subjected to Linear Discriminant Analysis (LDA), thus reducing the feature vector to represent the signature with optimal set of features. The merits of DCT that captures the significant information in a small pack of coefficients is fed into discriminant analysis for further compact representation. The proposed approach, DiscriminativeDCT - MLP combines the benefits of two domains, yet does not suffer from their individual limitations. The optimal representative features from all the samples in the dataset form the knowledge base. Further, the Multi-layer perceptrons (MLP), a well known classifier is used for classification and the performance is measured through FAR/FRR metrics. Experiments have been conducted on standard signature datasets namely: CEDAR and GPDS-160, and MUKOS, a regional language (Kannada) dataset. The comparative study is also provided with the well known approaches to exhibit the performance of the proposed approach.