Improved writer retrieval in handwritten documents using hybrid combination

Writer retrieval has valuable applications in analyzing handwritten documents, such as the verification of authenticity and authorship of unknown manuscripts. Writer retrieval systems are designed to automatically locate all manuscripts attributed to a specific author based on dissimilarity measures between document features. In the present work, we propose a hybrid combination of handcrafted and deep features to achieve robust writer retrieval. Two writer retrieval systems based on deep features are implemented using the VGG-16 and the MobileNetV2 models. Then, in order to bring complementary information, we propose the MO-HOT (Multi Orientated Histogram Of Templates) to develop shape features-based writer retrieval system. The MO-HOT highlights the writing traits in various directions along with various scales. The three systems are combined through a SVM based dissimilarity learning to aggregate a final writer retrieval decisions. Experiments are conducted on four handwritten document datasets that are CVL, ICDAR-2011, ICDAR-2013, and ICDAR-2017. The results obtained evince the effectiveness of the proposed hybrid combination, which outperforms the best state of the art results with up to 2% in the TOP-2 score. Besides, the proposed MO-HOT provides comparable and complementary performance with deep features since it helps to improve the MAP score by at least 3.6% for CVL, ICDAR-2013 and ICDAR-2017 datasets.