Dual-Path JND: A New Framework for Robust and Imperceptible Image Watermarking

The Just Noticeable Difference (JND) model is widely used in image watermarking to balance imperceptibility and robustness by aligning embedding strength with human perception. However, conventional JND models based only on low-level pixel statistics often misallocate embedding energy, over-embedding in sensitive regions while underusing tolerant areas. We propose Dual-Path JND, a framework that integrates image-driven cues with deep perceptual priors for context-aware modeling. Besides the pixel-based branch, a deep-prior branch extracts and fuses multi-scale features from a pre-trained backbone to generate a perceptual guidance map capturing fine textures and high-level semantics. Both branches generate complementary JND estimates through JND modeling processes, which are then fused into a content-adaptive embedding strength for intelligent watermark allocation. Experiments on COCO, LSUN, and DIV2K show that our method achieves a superior imperceptibility–robustness trade-off, with up to +2.0 dB improvement in peak signal-to-noise ratio and consistent gains in structural similarity index, while maintaining high detection accuracy (true positive rate ≥99.69%, bit error rate ≤0.60) under diverse distortions.