MCFN: Multi-scale Crossover Feed-forward Network for high performance watermarking

In recent years, deep learning-based watermarking for digital images has seen substantial progress, particularly in embedding secret messages in an invisible manner and extracting them after attacks. However, existing methods face challenges in effectively integrating spatial and frequency domain features, which limits overall performance. To address this, we propose the Multi-scale Crossover Feed-forward Network (MCFN), an innovative solution explicitly designed to process features of secret messages and cover images in a robust watermarking framework. The MCFN features a crossover structure that facilitates global communication between single-dimensional and double-dimensional features, enhancing the capacity to simultaneously capture crucial characteristics from both secret messages and cover images. A key element of our design is the Discrete Wavelet Transform (DWT) convolution block, which enables multi-scale feature extraction in the frequency domain, overcoming the limitations of conventional watermarking networks. We also incorporate perceptual loss to ensure high visual quality of the watermarked image and introduce a Bi-Directional Attack Simulation Layer (BD-ASL) to improve resilience against various attacks during training. Extensive experimental results validate that our framework achieves superior capacity, robustness, and imperceptibility, significantly outperforming current state-of-the-art methods in the deep learning-based watermarking field.