Research Of Source Printer Authentication For Printed Documents Based On Texture Features

Increasing demand of information security calls for new printed document examination technology to identify the source printers of the questioned documents, which is also named Source Printer Authentication (SPA). Printers are manufactured with large-scale standardization, and printed documents are standard and convergence, which brings forth a higher technology challenge for SPA. There currently exist techniques to secure documents such as bank notes using paper watermarks, security fibers or special inks. The problem is that the use of these security techniques can be cost prohibitive. Most of these techniques either require special equipment to embed the security features, or are too expensive for an average consumer. It’s desirable to be able to identify the manufacturer, or model that was used to print a given document even if the printer in question does not make use of these existing security devices to explicitly identify itself. It would be useful both in theory and practice to achieve the same or better level of protection without the use of any additional devices or expensive technologies.Under this background, current SPA methods are investigated and concluded, the difficulties of SPA are analyzed and practicable scheme based on texture features are provided in this thesis. The main contents and innovations of this thesis are listed as follows.1) The texture differences of printed document images from different printed are further investigated under the condition of high magnification. Two kinds of texture features are implemented for printed character images feature extraction, one is Gray level Co-occurrence Matrix(GLCM) method, which is based on statistics, the other is Wavelet Transformation(WT) method, which is based on digital transformation field. Support Vector Machine (SVM) is then used to classify the printed documents and high correct rates are achieved. The experimental results illustrate that the textures features of printed document images are effective for SPA.2) Factor variance analysis model about printer texture element is built to examine the significance of printer texture elements, and the experimental results prove that texture information can be the foundation of SPA. Then two-way factor analysis model of texture factor and character factor is built to analyze and examine the influence of these two factors for printed document images. Finally, the mathematical fundamental model for SPA is built and it proves both in theory and experiment that the correct rate of SPA can achieve 100% based on texture features when some conditions are satisfied.3) Current SPA methods are invalid when there are no identical characters between the training and testing documents, In order to address this problem, a printed documents recognition method based on texture filling is proposed. The character images are seen as the combination model of texture and character structure. Then the improved Criminisi algorithm is explored to fill the blank space and obtain completed texture image. Thus the character structure is erased and the new texture images are character-independent. GLCM and WT methods are used for feature extraction and SVM is used for classification. The experimental results illustrate that this proposed method overcomes the disadvantages of traditional texture methods and broaden the usage range of texture features.4) The SPA method based on Gaussian Mixture Model is proposed. It can be seen from the analysis of printed character images that the image texture doesn’t obey the single Gaussian distribution, but can be divided into three parts:texture region, bound region and background region. Each part is analyzed respectively and approximately obeys Gaussian distribution. Then a character image is modeled as mixture of three Gaussian models. K-means algorithm is used to find appropriate initial value of model parameters. Final parameters of the models are derived using Expectation Maximization(EM) algorithm. The model parameters are then used as texture features to train SVM and experimental results are satisfied.The SPA problem is further investigated and the texture information of printed documents are researched and discussed in this thesis. Variances of texture information among printed documents from different printers are fully considered and effective texture features are proposed for SPA. The experimental results are promising and improved substantially compared to previous methods. The SPA technique is more applicable than ever.