Rethinking image watermarking for better embedding

Image watermarking plays a critical role in protecting intellectual property. While most state-of-the-art (SOTA) methods based on deep learning enhance watermark performance, they often fail to consider the variability of embedding differences among different cover images. This oversight neglects prior image knowledge, which limits embedding performance. To address this issue, by associating the significant spatial redundancy in natural images that can well depicts the prior image knowledge, we propose a novel local watermarking framework based on <i>local consistency</i>, called <strong>LocMark</strong>. Unlike traditional digital watermarking methods that target texture-rich regions, <strong>LocMark</strong> adaptively embeds watermarks into highly consistent smooth regions. Specifically, to identify these consistent regions, we design a simple yet effective region localization mechanism by correlating local consistency with pixel distribution. By normalizing the total gradient to the scale of one region, a variable-length training strategy based on gradient accumulation is further proposed to handle varying regions. Compared to the baseline methods, <strong>LocMark</strong> achieves SOTA performance with an average PSNR increase of 2.96 dB and a significant reduction in detection rate by steganalysis of 32.85%.