A Multi-patch Deep Learning System for Text-Independent Writer Identification

Abstract

Offline handwriting identification is widely used in many fields of modern society, such as judicial authentication, identity verification, ancient manuscripts, etc. Compared with traditional methods, methods based on deep learning have been proven to extract more distinctive handwriting features from large amounts of data and show better performance. For deep learning methods, data preprocessing and global feature coding determine its performance. To solve this problem, this paper proposes an offline writer identification method, which combines multi-patch data preprocessing and transfer learning. First, multi-patching and data enhancement techniques are used to process handwritten images. Subsequently, the pretrained residual network of the image dataset is used for local feature extraction, and finally, the mean feature method is used for global feature encoding. The proposed system is tested on ICDAR2013 and CVL standard datasets. Experimental results show that this novel method has a relatively stable and good recognition effect