Text-independent Writer Recognition Using Edge Co-occurrence Feature With The Aid Of Character Pairs

Writer recognition is one of hot topics in the research areas of handwritten document analysis and biometrics.It raises concern from both academia and industry for a long time.Writer recognition can be widely applied in many fields,such as finance,security,forensics,and so on.Text-independent writer recognition is to identify a person on the basis of handwriting using features designed to be independent of the textual content of handwritten samples.In the past two decades,lots of writer recognition methods have been developed.But due to the complexity of practical applications,the exsiting writer recognition methods are still not perfect.This thesis focuses on the off-line text text-independent writer recognition methods,and investigates the related issues of feature extraction,feature combination and similarity fusion.We propose four distinctive writer recognition methods,and validate their effectiveness by experiments.The contributions of this thesis are summarized as follows:1.An edge co-occurrence feature,which is a histogram-based feature,is proposed for text-independent writer recognition.The proposed edge co-occurrence feature treats the co-occurrence edge pixel pair of the each edge pixel as the elemental statistical unit,and characterizes the handwriting by the distribution of co-occurrence edge pixel pairs,which takes into account not only the directions but also the positions of the edge pixels.The proposed feature achieves satisfactory experimental results on several public datasets and competition datasets,which demonstrates that the proposed feature is effective for distinguishing the handwritings with different writing styles.2.A text-independent writer recognition method using the edge co-occurrence feature and histogram of SIFT descriptors is proposed.The histogram of SIFT descriptors records the textural information of keypoints extracted from the scale-space difference of Gaussian images.At first,the SIFT descriptors extracted from training handwriting images are used to create the codebook by a clustering algorithm.Then,with the codebook,the histogram of SIFT descriptors of the test handwriting image are calculated.Compared with the edge co-occurrence feature,the histogram of SIFT descriptors describes the handwriting from the macro-perspective.Thus,the combination of the edge co-occurrence feature and the histogram of SIFT descriptors can improve the performance of the writer recognition system.Experimental results on several public databases and competition datasets validate the effectiveness of the proposed method.3.A Chinese writer recognition method with the aid of the similarity of characters with the same text content appearing in the query and reference handwriting is proposed.After optical character recognition,we employ SIFT descriptor to calculate the similarity between two character images with the same text content.The character pair similarity is employed to assist the edge co-occurrence feature in writer recognition.Experimental results on two Chinese public databases validate the effectiveness of the proposed method.4.A Chinese writer recognition method with the aid of character pairs detected by the displacement field is proposed.The proposed method is independent of Chinese character recognition.The displacement field of two character images is utilized to calculate the character pair similarity.The displacement field based character pair similarity is used to help the text-independent similarity for writer recognition.Experimental results demonstrate that our proposed method outperforms the state-of-the-art methods of Chinese writer recognition.In addition,the proposed method also improves the performance on the English dataset,which implies that the proposed method can be extended to other languages.