Research On Robust Text Image Authentication

Application of robust text image authentication originates from the protection of printed documents. After digital documents go through print-scan or other common operations, the watermark should be still extracted perfectly to guarantee the security of authentication. These algorithms have to be robust enough to geometrical distortions and quantized noise. The basic idea of digital watermarking is taking advantage of the redundant information of multimedia data. However, little redundancy can be employed in text images in contrast with other multimedia signals, which brings great difficulties for robust text image authentication and then makes it hard to develop.In multimedia content authentication, the focus of current researches are mainly on the authentication watermarking and digital signatures. Experience has proved that the safest way is to use the signatures of original images as message authentication codes to be hidden into the images for secure authentication watermarking. Most existing semifragile watermarking techniques applied to robust authentication suffer from poor robustness and low capacity. As an emerging research direction, robust signatures developed today are mainly for continuous-tone images, video and audio, but rarely mentioned in text image authentication. This paper analyzes the general framework of robust text image authentication, and works on the above two issues. The main contributions are presented as follows.(1) A semifragile text image watermarking scheme based on stroke direction modulation for Chinese documents is proposed. The angles of rotatable strokes are quantized to embed the bits. This requires several stages of preprocessing, including stroke generation, junction searching, and rotatable stroke decision. Considering the visual masking and security properties, the shuffling and neighborhood similarity constraints of embeddable strokes are employed. During the detection phase, the compensation measure of channel distortions based on Hough transform is introduced. Compared with some existing methods, the proposed method achieves more favorable rate-distortion-robustness trade-offs.(2) A contour feature-based text image watermarking scheme against print and scan processes is proposed. Based on the print-scan invariant, the boundary points of each character are flipped using Fourier descriptors with visual perception identity, so that the marks are embedded into the visually nonsignificant points. In the calculation process of the print-scan invariant, a certain text line serves as the reference line to compute the population mean of characters’ pixels, for not affording additional characters to keep the sum of black pixels consistent. Thus, the hiding capacity is greatly improved. The algorithm is not limited to a particular language, and has better self-adaptability, watermark transparency as well as hiding capacity compared with some existing methods.(3) A character structure based approach to robust text image hashing is proposed. Two types of low-level features are developed in the algorithm, viz, stroke-based and junction-based, to eliminate the perceptual redundance. Then the cryptographic hashing is used for further compressing the feature vector. Some compensation measures have to be taken into account for the distortions induced by hardcopy operations such as baseline skewing. The generated hashes have good perceptual robustness, but also have the properties of one-wayness, randomicity and collision resistance, which ordinarily characterize a cryptographic hash. The effectiveness of the proposed algorithm under hybrid electronic-analog channels is demonstrated by the experimental results.(4) A character shape based approach to perceptual text image hashing is proposed. Fourier transform features and skeleton-based invariants are combined to produce an intermediate hash. The final hash is obtained by compressing the quantized hash vector using non-negative matrix factorization (NMF) and pseudorandom statistics. Unlike robust image hashing, an ideal perceptual image hash is a good measure of the perceptual distance between two images, which can better represent the perceptual content by human visual system. Three independent keys in different stages, i.e. feature extraction, quantization and compression, confer excellent security on the algorithm.