Offline signature verification is a challenging problem in the field of biometric recognition due to the presence of a lot of intra-writer variations and skilled forgeries. Conventional approaches in signature verification using only symbolic features and a predefined measure of similarity often fail to handle the complexities of visual patterns. To overcome this problem, we propose a new framework for signature verification using a combination of symbolic and convolutional neural network features. The symbolic features are able to capture the structural and statistical properties of signatures, whereas the convolutional neural network features are able to capture the high-level spatial features of signatures. The proposed framework uses a weighted feature-level fusion of both representations. The proposed framework achieves better performance in the presence of limited samples using the cosine similarity function and optimizes the decision threshold using the average error rate.
Manasa et al. (Thu,) studied this question.