Score Level Fusion for Improving Writer Retrieval in Handwritten Document Databases

Writer retrieval in handwritten document databases is used in various aims such as authenticating old manuscripts, and predicting the writer of anonymous texts. Such systems try to retrieve all documents of a given writer without having a prior knowledge about its identity. This paper introduces the use of score level fusion for improving the retrieval results. Several rules that are based on Sum operator, triangular Norms and Co-norms are employed to combine outputs of several writer retrieval systems. The fusion step is carried out by considering three individual writer retrieval systems, that are based on Edge Hinge Distribution features, Run-length Local Binary Pattern and Histogram Of Templates. Experiments are conducted on ICDAR2011 dataset. The results obtained reveal that fusion rules can give an improvement around 2% in MAP over the scores derived from individual systems.