Writer Identification Based On Local Wavelet Binary Patterns

Writer identification is widely used in forensic, commercial applications and histori-cal documents analysis. Automatic writer identification systems using computer technologyhave received an increasing interests and become an active research topic. Conventionalwriter identification algorithms usually work on the binarized handwriting images whichlose some useful writing style information (e.g. gray level intensity) and require constrainedwriting contents. Local descriptors, which extract image features from constrained supportareas, have be successfully applied to many computer vision problems (e.g. object detec-tion, action recognition, texture analysis ect.). In this paper, we cope the writer identificationproblem following the framework of local feature histogram which assumes that handwrit-ing images are different kinds of textures. This allows us borrow some methods and toolsfrom texture analysis which has been intensively researched. The performances of differentlocal descriptors are illustrated and tested. In addition, we proposed new local descrip-tor , local wavelet binary patterns (LWBP). Experiments validate the effectiveness of theproposed method. Compared with other algorithms, it has desirable properties like higherdiscriminative ability, more computational efficient and robust to noise. Since the proposedmethod is based on non-parametrical histogram comparison, it has no assumptions aboutthe distribution of the features and does not need tuning of the parameters. This preventspossible wrong assumptions of feature model. Also, it does not require complex learningprocedure.The main contributions of this thesis is listed below:(1)The background, applications and some conventional methods of writer identifica-tion are illustrated. In addition, the analysis of conventional writer identification methods ispresented.(2)Different kinds of local descriptors are introduced and compared. This thesis pro-pose are general framework for future writer identification research based on feature his-togram.(3)The proposed new local descriptor , local wavelet binary patterns (LWBP) is illus-trated. Based on this new feature, a writer identification algorithm is developed. Experi-ments prove that the new feature is effective in extracting writing style information.(4) The conclusion and the prospect are given at the end.