Offline Signature Verification with VLAD Using Fused KAZE Features from Foreground and Background Signature Images

Offline signature verification has been widely accepted as a tool for individual authentication, especially in the field of biometrics and forensics. However, the performance of an automated system under a wide range of writing conditions is still inadequate. The promising approach is to consider recent knowledge about the cognitive processing of visual information of forensic document examiners (FDEs). To implement FDEs' cognitive processing method successfully into offline signature verification, specifically this study proposes a new approach based on vector of locally aggregated descriptors (VLAD) with fused KAZE features detected from foreground and background signature images with a recent fusion strategy. The experimental results by the proposed method with a popular MCYT-75 signature dataset can be summarized as follows: (1) the KAZE features from the background signature images as well as the ones from the foreground images show good performance. (2) The use of fused KAZE features from foreground and background signature images allows us to further improve of the performance. (3) Among the typical fusion methods, the representation-level fusion is a rational choice for fusing the KAZE features to obtain good performance. (4) While the representation-level fusion produces a high-dimensional VLAD vector, the use of principal component analysis for the original VLAD vector can provide a more dimensionally compact vector without significant loss in performance. (5) Finally, the proposed method provides much lower error rates than the existing state-of-the-art offline signature verification methods.