Dual-path watermark embedding model based on generative adversarial network

Abstract Digital watermarking technology based on deep learning has the following drawbacks: watermark embedding tends to cause visual distortion of the host image; feature loss of extracted watermarks is significant under different attack types and intensities; and the balance between watermark invisibility and robustness is limited. To tackle these problems, this paper puts forward a dual-path fusion watermark embedding model relying on generative adversarial network (DFEGAN). First, a dual-path watermark embedding method is proposed for the embedding network, which weight-fuses the embedded images derived from the dense attention network (DAN) and discrete wavelet transform (DWT) paths. This method effectively mitigates host image distortion induced by watermark embedding. Second, a global channel aware attention module (GAM) is designed for the extraction network, where the attention weights computed by this module enhance the channel features of watermarked images. Meanwhile, a multi-scale convolution fusion module (MFM) is proposed. It captures and integrates large, medium and small-scale watermark features (including contours, textures and brightness) using convolution kernels of different sizes, yielding multi-dimensional watermark features. This design effectively enhances the watermark’s resistance against attacks of varying types and intensities. The results demonstrate that DFEGAN outperforms the comparative models in the experiments.