Multi-Path Siamese Convolution Network for Offline Handwritten Signature Verification.

The existing methods are difficult to maintain the balance of high accuracy and low model capacity for writer-independent offline signature verification. Thus, this paper developed a novel deep learning model named Multi-Path Attention Siamese Convolution Network (MA-SCN), which consists of three modules: the generation module of different size inputs, ensuring the same size of the signature strokes and the different size of the image background; the feature extraction module based on the fusion attention mechanism, focusing on the stroke features; the verification module with optimal weight coefficients, improving the model's performance. The experimental results show that the setting of different factors such as the multi-branch structure, the fusion attention mechanism, and the weight coefficients is effective for improving the accuracy of offline signature verification on the three public datasets (CEDAR, Bengali, Hindi). Moreover, compared with the state-of-the-art offline signature verification model, our model has fewer parameters.