Owing to the rapid growth of touchscreen mobile terminals and pen-basedinterfaces, handwriting-based writer identification systems are attractingincreasing attention for personal authentication, digital forensics, and otherapplications. However, most studies on writer identification have not beensatisfying because of the insufficiency of data and difficulty of designinggood features under various conditions of handwritings. Hence, we introduce anend-to-end system, namely DeepWriterID, employed a deep convolutional neuralnetwork (CNN) to address these problems. A key feature of DeepWriterID is a newmethod we are proposing, called DropSegment. It designs to achieve dataaugmentation and improve the generalized applicability of CNN. For sufficientfeature representation, we further introduce path signature feature maps toimprove performance. Experiments were conducted on the NLPR handwritingdatabase. Even though we only use pen-position information in the pen-downstate of the given handwriting samples, we achieved new state-of-the-artidentification rates of 95.72% for Chinese text and 98.51% for English text.
展开▼
