Off-Line Signature Verification using Novel Feature Extraction Techniques and Trajectory Recovery

Abstract

Feature extraction is an important process in automatic off-line signature verification systems. In this process, only the information that helps to identify the authenticity of questioned signatures is extracted and retained. Amongst the numerous feature extraction techniques investigated by researchers, grid segmentation schemes have been employed more favourably due to their encouraging results. The research presented in this dissertation focuses on improving the performance of Support Vector Machines (SVMs) based off-line signature verification systems using novel feature extraction techniques. The research began with an in-depth investigation and comparative performance analysis of the Modified Direction Feature (MDF), a structural feature extraction technique. Since the MDF is known for its relatively high accuracies in the cursive character recognition problem, it has been suggested that the performance of the MDF would be as encouraging in verifying Western (English) signatures due to the “cursive appearance” of the signatures. The other features employed for comparative studies were the two grid-based features proposed by Francesco Camastra and Wakabayashi et al. (Gradient feature). The former feature captures the information about the distribution of the foreground pixels in each grid cell whilst the latter utilises the directional information available. The comparisons of these state-of-the-art techniques set the foundation for the development of the novel feature extraction techniques proposed. In total, three novel local features and four global features were proposed and investigated. The local features include Gaussian Grid, Curvature Map, Variance and the local features are New Ratio, Energy, Trajectory Length, Moment. Amongst the local features proposed, the Gaussian Grid feature significantly outperformed all the state-of-the-art features mentioned above. Nevertheless, the combination of another particularly small-dimensional local feature, the Variance feature, with the global features also outperformed the MDF and the Camastra features, and closely approximates the performance of the Gradient feature. The total dimension of this feature set was only 33 compared to 120 of the MDF. This finding emphasizes the capability of small-dimensional global features.