Normalized Intrinsic Deep Features Based Zero-Watermarking Scheme for Remote Sensing Images Using U-Net and K-Means

Most remote sensing image zero-watermarking algorithms are designed for specific types of data and particular attacks. They overly rely on manually designed features, resulting in insufficient security, poor generalization, and inadequate robustness. To address the issue, a novel zero-watermarking algorithm based on U-Net and K-means is proposed. The algorithm employs a neural network to extract intrinsic deep features from remote sensing images to construct zero-watermarks, thereby enhancing the algorithm’s security while ensuring adaptability to different types of remote sensing images. The algorithm begins by selecting scale invariant feature transform (SIFT) feature points within a specific intensity range, constructing a Delaunay triangulation, and using its incircles as feature domains. During feature extraction, a specially trained U-Net network is employed to extract robust features for generalized recognition across multitype images. While constructing the feature matrix, robust features are clustered using K-means, and orientation information is normalized to ensure high uniqueness of the generated watermark. The experiments demonstrate that the algorithm exhibits excellent uniqueness and highly resistance to conventional image processing (e.g., noise, filtering and compression), geometric attacks (e.g., affine transformations and cropping), and combination attacks. Additionally, it has good applicability to remote sensing images from various sensors, precision levels, and spectral bands.