Application and research of image watermark based on free visual size optimization JND algorithm

To address the problems of perceptual redundancy estimation in digital image watermarking, which are poor adaptability to fixed-size inputs and insufficient coordination between local and global visual features, a perceptual redundancy estimation based on perceptually sensitive size optimization is proposed. First, BI and block aggregation are combined to achieve size adaptation of visual saliency maps. A multi-scale saliency map fusion mechanism is then used to combine local details with global structural information. An adaptive weighting strategy based on edge complexity is introduced to dynamically modulate feature contributions. Finally, the optimized model is integrated into an extended transform dithering modulation watermarking framework to achieve perceptually adaptive embedding. The proposed method is validated on the USC-SIPI and BSDS300 datasets. Experimental results show that the proposed method improves structural similarity and peak signal-to-noise ratio by 5.38% and 2.67%, respectively, and improves the normalized correlation coefficient by an average of 4.05% when resistant to compression attacks. The multi-scale fusion mechanism improves the normalized correlation coefficient by 5.96% and reduces the bit error rate by 18.97% under noise and filtering attacks, respectively. Compared with mainstream methods, this model performs best in both compression and noise resistance, achieving a maximum increase in the normalized correlation coefficient of 7.08% and a maximum reduction in the bit error rate of 34.78%. It also maintains a stable lead in all objective indicators across both datasets. The results demonstrate that the proposed method can effectively improve the imperceptibility and robustness of watermarks, providing a more adaptable and high-performance solution for image copyright protection. The core contribution of this paper lies in systematically addressing the two major bottlenecks of the traditional JND model when applied to watermarks: by using a scale-adaptive mechanism, the dependence on a fixed input size was overcome; through multi-scale fusion and content-adaptive weighting, the collaborative optimization of local details and global structure perception was achieved, thereby more accurately depicting the visual masking effect in complex texture regions.