KAZE features via fisher vector encoding for offline signature verification

The widespread use of handwritten signatures for identity authentication has resulted in a need for automated verification systems. However, there is still significant room for improvement in the performance of these automated systems when compared with the performance of human analysts, particularly forensic document examiners, under a wide range of conditions. Furthermore, even with recent techniques, obtaining as much information as possible from a limited number of samples still remains challenging. In this study, to tackle these challenges and to boost the discriminative power of offline signature verification, a new method using KAZE features based on the recent Fisher vector (FV) encoding is proposed. The adoption of a probabilistic visual vocabulary and higher-order statistics, both of which can encode detailed information about the distribution of KAZE features, provides us with a more precise spatial distribution of the characteristics for a writer. The experimental results on the public MCYT-75 dataset can be summarized as follows: 1) The proposed method improves performance compared to the recent vector of locally aggregated descriptors (VLAD)-based approach. 2) The use of principal component analysis (PCA)for the original FV can provide a more dimensionally compact vector without a significant loss in performance. 3) The proposed method provides much lower error rates than existing state-of-the-art offline signature verification systems when applied to the MCYT-75 dataset.