Résumé: Since diffusion-weighted imaging (DWI) images are high-dimensional medical images with rich texture features, existing traditional zero-watermarking algorithms cannot effectively protect the copyright of DWI images while satisfying the conditions of equilibrium and distinguishability. In this paper, a robust zero-watermarking algorithm for DWI images based on multiscale feature fusion is proposed to achieve effective copyright protection of DWI images. Firstly, a Siamese network for multiscale feature extraction based on the imaging characteristics of DWI images is proposed to extract robust features of two-dimensional slice sequences in each diffusion gradient direction by using different transforms. Next, to address the low distinguishability of extracted features due to the similar anatomical structure of different people under the same medical imaging mode, a variety of features based on texture and spatial correlations are integrated as a prior knowledge features into the network to enhance the distinguishability of zero-watermarking. Then, the dimensionality of the extracted features is reduced by using a clustering algorithm, and the key slice images are selected in each category of each diffusion gradient direction by using the clustering results. The eigenvalues of the eigenmatrix of the key slices are extracted by singular value decomposition to generate a binary feature map according to the gray average of the eigenvalue matrix. Finally, the binary feature map and the watermarking signals are modulated by logic chaos, and the zero-watermarking information is generated by performing the XOR operation. The experimental results show that the proposed algorithm is robust against various intentional or unintentional attacks on medical image processing, such as noise, filtering, JPEG compression and geometric attacks. The extraction accuracy rate of watermarking is above 97%, and it is more robust than traditional robust zero-watermarking methods.

Langue: Anglais