A Handwritten Signature Segmentation Approach for Multi-resolution and Complex Documents Acquired by Multiple Sources

Handwritten Signature is a biometric feature, which enables personal verification. Thus, it constitutes an alternative authentication used in several applications, such as bank checks, contracts, certificates, and forensic science. Signatures may be presented on a complex background with different textures, turning automatic signature segmentation into a difficult task. In this work, we propose an approach to locate and segment only the signature image pixels in documents with complex backgrounds, acquired by smartphone cameras in different environments, without any prior information about the signature location in these documents, the pen used by the signer, among others issues. Our approach is based on the U-net network architecture, combined with a pre-processing stage that allows dealing with images having different resolutions and distortions due to the document acquiring process. To make our model more robust to background and texture variations, we have generated a data set consisting of 20,000 document photos with different sizes, textures, and documents, named DSSigDataset-2. Our experiments show that the proposed method achieved encouraging results, over precision, recall, and F1-score measures, in all evaluated data sets (ours and benchmark ones).