Offline Writer Identification Using Convolutional Neural Network and VLAD Descriptors

Abstract

This paper proposes a method for writer identification and retrieval. The focus of these two tasks is to extract discriminative features from handwriting documents. Traditional artificial features require researchers’ experience and professional knowledge in language and handwriting, while the automatic features are extracted from large data handwriting samples. The proposed method uses local features and has better performance. We extract features via the convolutional neural network, subsequently aggregating to obtain the global feature vector. During the process, the system first enhances and preprocesses the handwriting samples and then feeds the patches into the ResNet50 that pretrained on ImageNet. Finally, the local features, i.e., the output of ResNet50, are aggregated into a global feature vector by the VLAD method. The evaluation is employed on the ICDAR2013 and CVL data sets. Experiments show this approach is close to the state-of-the-art methods, and it has good stability and robustness.