Writer Identification with Hybrid Edge Directional Features and Nonlinear Dimension Reduction

Writer identification as an interesting pattern recognition topic has attracted many researchers in forensic and biometric applications, where the identity authentication is realized by the writing style as biometric features. The core issue in writer identification, therefore, is the extraction of unique features, by which the individualistic of such handwriting styles can be faithfully represented. Many writer identification methods based on handwritten text lines have been proposed recently. In this paper, an effective method to perform writer identification is proposed by extracting edge directional features based on the recently proposed local shape descriptor Pyramid Histogram of Oriented Gradients (PHOG) and a curve fragment feature which can be termed as Histogram of Edge Direction (HED). PHOG represents the local shape by a histogram of edge orientations computed for each image sub-region, quantized into a number of bins. HED is based on a new edge detection and edge linking algorithms which robustly extract well-localized sub-pixel edges and stably links these into curve fragments. Compared with many previously proposed feature extraction approaches in writer identification research, PHOG and HED have the advantages in the effective extraction of discriminating information from handwritten text images and a hybrid feature from the combination of them can give high identification accuracy. The identification performance can be further enhanced by applying a graph embedding based nonlinear dimension reduction. The writer identification experiment is thoroughly studied in the general pattern classification framework, with five different commonly applied classifiers compared, including linear discriminant analysis (LDA), Naive Bayes, k-Nearest Neighbor, multiple layer perceptron (MLP) and Support Vector Machine (SVM). The experiments Preprint submitted to ??? January 28, 2010 have shown that the classifiers can all give very satisfactory performance but LDA is the best. On a public benchmarking IAM database, experiments demonstrates an average accuracy 96% from LDA, which compares sharply with the latest published results on the same dataset, thus illustrating the superiority of the proposed method.