Calligraphy Writer Identification Based On Gabor Filter And Gaussian Markov Random Field

This article assists to identify the authenticity of calligraphy works from the perspective of handwriting analysis. Although currently there is very little study on calligraphy writer identification, the off-line text-independent Chinese character handwriting identification methods provide a great important reference for our research on calligraphy writer identification. Extracting effective features to describe handwriting is always a key problem in writer identification. This paper mainly studies handwriting feature extraction algorithm based on texture analysis. The main contents of the dissertation are as following:Firstly, we provide a brief review of the relevant background of calligraphy handwriting identification and the research status about the feature extraction methods of off-line, text-independent writer identification.Secondly, we complete the preprocessing of original calligraphy images and propose to use histogram of gradient(HOG) of the character strokes to optimize the orientations of Gabor filter. In order to overcome the shortcomings of the traditional Gabor filter method, as well as to fully exploit correlation between Gabor filtering coefficient. We propose to apply Gauss Markov Random Field(GMRF) model on every filtered image to describe the different local spatial structures, and successfully merges the global and local features together.Finally, we propose a novel method for writer identification combining Gabor filter and Gaussian Markov Random Field(GMRF). The basic idea is: firstly use the optimized Gabor filter to extract texture features and singular information of the handwriting image, which highlight the overall features of handwriting in a certain direction and spatial frequency; Then apply the Gaussian Markov Random Field model to describe hidden local structure of the filtering image, analyzing local structure information of handwriting. The feature extraction algorithm considers both micro-structure of handwriting and overall writing style. With the four most famous regular script writers’ original samples and the collected English scripts as the experimental data, the minimum weighted Euclidean distance classifier is applied to classify handwriting samples, respectively achieving correct classification rates of 93.3% and 87.4% with five-fold cross validation method. Our experiments show that, compared to the traditional Gabor filter method or the single Gaussian Markov Random Field model method, the handwriting features extracted with the combination of the Gabor filter and GMRF method we proposed have stronger ability to characterize the handwriting, and the combination method gets a more satisfied result in the handwriting sample pool used in this dissertation.