Research On Digital Watermarking Technique Based On Image Local Features

The popularity of Internet access and rapid development of the multimedia techniques make copyright protection of digital multimedia facing a new challenge. For this reason, digital wa-termarking, as an emerging technology to protect digital multimedia from unauthorized redistri-bution, becomes a research topic. Attacks have been developed to destroy watermark for water-marking is used to copyright protection. In this case the watermarking approaches used in copy-right protection should be robust to these attacks. In recent years, robust watermarking research has made great progress and many robust watermarking algorithms have been proposed. How-ever the robustness of these algorithms, especially the robustness to geometrical attacks needs to be improved. This thesis discusses watermarking based on local image features and makes some progresses. The main work of this thesis is as the following:1. Based on the stability and invariance of the local feature points, this thesis proposes re-synchronization and self-synchronization robust watermarking framework. The resynchroniza-tion watermarking framework corrects synchronization errors using feature matching and image geometric correction. Many existing embedding and extracting methods can be used in this framework. But it should be noted that geometrical attacks and correction in resynchronization induce some interpolation errors for watermarked image, which will decrease the accurate rate of the extracted watermark. The interpolation errors are related to the local variance of local pixel gray values. Changing the watermark embedding intensity dynamically based on human visual system can reduce the adverse effect of the interpolation errors. Self-synchronization watermark-ing framework extracts feature points from carrier image, and chooses some stable points to mark some local regions. The copy of watermark is embedded into every region repeatedly. The watermark extracting algorithm detects some feature points from the watermarked image and searches the watermarked region. The Self-synchronization of watermark is kept by the invari-ance of feature points, local feature region and watermarking method.2. An image geometric correction method is proposed based on Scale invariant Feature Transform (SIFT). The whole process of this method includes SIFT feature extracting, feature matching and geometric attacked image correction. This thesis analyzes SIFT schemes, and gives an improving method which applies SIFT local features in the resynchronization water-marking framework. Firstly it improves running speed by optimizing the convolution method for Gaussian Scale Space. Second it improves feature detecting method to get smaller number of stable features, which can get description and matching efficiency. Lastly it changes unidirec-tional matching strategy to bidirectional matching in order to get better precision. Utilizing the coordinates of the matched feature points it can calculate geometric transformation parameters. With these parameters it performs geometric anti-transformation to correct the attacked image by geometric transform. Experimental results show that the algorithm is effective to rotation, scale, and slight affine transformation.3. This thesis researches the quantization index schemes based on image geometric correc-tion, and gives three adaptive quantization index watermarking schemes in spatial domain, DCT domain and DWT domain respectively. It firstly analyzes distribution of the quantization errors, and give the relationship of PSNR, watermark capacity and quantization step in spatial domain, DCT domain and DWT domain respectively, then decides the basic quantization step using the method. The proposed watermarking schemes adjust the basic quantization step dynamically ac-cording human visual properties. In order to strength the robustness a watermark bit is embedded at pixels or transforming domain coefficients repeatedly. The watermark bit is extracted accord-ing to the number of extracted'0'or'1'. Experiment results show that adjusting the quantization step dynamically according human visual properties not only can reduce the effects of the inter-polation errors induced by geometrical transform, but also is robust to changes in amplitude and Jpeg compression.4. A spatial domain adaptive watermarking scheme is presented based on the visual mask of local image regions and correlation of the adjacent image pixels. It uses fuzzy c-mean clustering algorithm (FCM) to combine several visual mask features in spatial domain, in which two fuzzy groups are set, and the membership degree of the cluster with bigger visual mask feature value are regard as the factors of the embedding strength. It employs Neural Network to learn the rela-tion among the adjacent image pixels, and then adjusts pixel gray values to insert watermarks based on the outputs of the trained neural network, the membership degree of cluster and the embedding watermark bits. Experiment results show that this scheme is robust to the common image-processing operations. It can also be robust to geometrical attacks with help of image cor-rection based on feature points.5. This thesis presents a robust image self-synchronization watermarking scheme utilizing the local feature points. This scheme adopts local feature points to locate some local quadrilateral regions that meet some requirements. These regions are used to embed watermark. In every re-gion it locates the embedding points using the projective invariance of the cross ratio of four col-linear points. In this case the self-synchronization between watermark and carrier can be kept because the feature points, local quadrilateral regions and watermark embedding points are in-variant to geometrical transform. This watermarking scheme can get better performance to local attacks because watermark is embedded in every local quadrilateral region repeatedly. Experi-mental results show that the proposed method can strongly resist geometrical attacks such as ro-tation, scaling, translation, cropping, row and column removal and so on. It can resist the com-pound attacks of these geometrical ones. Furthermore it can resist some image processing opera-tions.