Evaluation of Gradient Descriptors and Dissimilarity Learning for Writer Retrieval

In order to take advantage from collections of digitized handwritten documents, effective indexing and retrieval techniques are required. This work focuses on automatic writer retrieval, which is the task of finding in a dataset, all documents written by the same person. Contrary to conventional writer retrieval techniques that are based on dissimilarity measures, we propose to use the SVM classifier to perform the retrieval task. First, local gradient features are used to generate handwritten features. Then, dissimilarities calculated between intra-writer and inter-writer documents are used to train a SVM to allow an automatic retrieval of all the writers documents. Experiments are conducted on CVL and ICDAR 2011 datasets. The performance evaluation of the proposed system is carried out comparatively to the cosine similarity. Results obtained evince a significant improvement offered by SVM, which gives comparable and sometimes better scores than the state of the art.