SVM-based writer retrieval system in handwritten document images

Digital libraries include huge amount of information that are continuously increasing with the need of storing various kinds of handwritten documents such as administrative forms, and cultural heritage manuscripts. Therefore, new emerging techniques such as writer retrieval are used to facilitate information extraction from archived documents. Basically, a writer retrieval system is composed of two main steps that are feature generation and dissimilarity measure. To achieve a robust retrieval, we propose the use of an SVM classifier trained to automatically separate intra-writer features from inter-writer features. For feature generation, we investigate the effectiveness of the Histogram of Oriented Gradients, Gradient Local Binary Patterns, and Local Difference Features. Experiments are conducted on three benchmark datasets. The results obtained evince a satisfactory performance of SVM, which can give better results than the state of the art.