| abstract |
The invention discloses an offline handwritten signature segmentation method based on a dual-branch neural network, which relates to an electronic signature recognition technology, learns rough semantic information of a document image, divides the document image into regions, and segments the signature image based on an encoding-decoding branch network. Obtain the rough segmentation signature; infer the details of the blurred area on the edge of the signature, continue to perform feature coding, learn the boundary details around the signature, and obtain a finer segmentation signature; divide the feature categories according to the region, and adaptively assign the feature centers to make the image features belonging to the same category. Move closer to the given feature center; the signature space structure is captured, the adversarial loss function is generated, the authenticity of the finely segmented signature is evaluated, and the dual-branch network segmentation is fooled to obtain a higher-quality signature output. Small-sized electronic signatures can be accurately segmented and identified on arbitrary positions and complex backgrounds. It can be applied to electronic signature handwriting recognition in situations such as scene complexity and blurred background. |