什么是图片隐水印（频域暗码）？
图片隐水印是一种人眼不可见的信息嵌入技术，将暗码隐藏在图片的频域系数中。与常见的可见文字水印或半透明 Logo 水印不同，隐水印对图片外观几乎无影响，但程序可以从图片中提取出隐藏的信息，用于版权保护和溯源追踪。
为什么能抗压缩、抗截图、抗拍照？
本工具采用DCT（离散余弦变换）频域水印技术，核心原理：
- 频域嵌入：将图片分成 8×8 像素小块，对每块做 DCT 变换，将信息调制到中频系数上。人眼对中频变化不敏感，但 JPEG 压缩等有损操作也主要影响高频，因此中频区域的信息最稳定
- 扩频通信：借鉴 CDMA 扩频通信原理，用伪随机序列（PN 序列）对水印比特做扩频调制。单个比特的能量分散到多个频率分量上，即使部分频率被破坏，整体相关性仍可恢复
- 相关性检测：提取时用相同的 PN 序列做相关运算，只有与序列匹配的信号会产生强相关峰，随机噪声的平均相关性趋近零，因此即使图片经过压缩、裁剪、缩放等处理，仍能可靠检出
技术细节
- 变换域：8×8 DCT 块，与 JPEG 压缩标准对齐，使水印自然抵抗 JPEG 压缩
- 嵌入位置：优先选择中频系数（如 AC(3,1)、AC(1,3) 等），避开低频（影响视觉）和高频（易被压缩丢弃）
- 强度参数：控制嵌入能量大小。日常使用 20-40 即可，对抗极端压缩（社交媒体二次压缩）可提高到 50+
- 差异放大图：嵌入后与原图像素差异放大 20 倍显示，可见水印能量主要集中在中频纹理
典型使用场景
- 版权保护：在摄影作品、设计稿中嵌入作者名称或版权声明
- 溯源追踪：给不同接收者的副本嵌入不同编号，泄密时追溯来源
- 防伪验证：在证件照、票据图片中嵌入验证信息
- 数据确权：在 AI 训练数据集中嵌入标记，检测未授权使用
所有计算在浏览器本地完成，图片不会上传服务器，隐私安全。输出为无损 PNG 格式，最大限度保留暗码信息。
What is Invisible Watermark (Frequency-Domain Steganography)?
An invisible watermark is a technique that embeds hidden information into an image's frequency-domain coefficients, making it imperceptible to the human eye. Unlike visible text or logo watermarks, it has virtually no visual impact, yet programs can extract the hidden data for copyright protection and traceability.
Why is it resistant to compression, screenshots, and photography?
This tool uses DCT (Discrete Cosine Transform) frequency-domain watermarking. The core principles:
- Frequency-domain embedding: The image is divided into 8×8 pixel blocks, each transformed via DCT. Information is modulated onto mid-frequency coefficients. The human eye is insensitive to mid-frequency changes, and lossy operations like JPEG compression primarily affect high frequencies, making mid-frequency the most stable region
- Spread spectrum: Inspired by CDMA spread-spectrum communication, watermark bits are modulated with a pseudo-random noise (PN) sequence. Each bit's energy spreads across multiple frequency components — even if some frequencies are damaged, overall correlation can still recover the signal
- Correlation detection: Extraction uses the same PN sequence for correlation. Only matching signals produce strong correlation peaks; random noise averages to zero. This enables reliable detection even after compression, cropping, or scaling
Technical Details
- Transform domain: 8×8 DCT blocks, aligned with JPEG compression standard for natural resistance to JPEG compression
- Embedding positions: Mid-frequency coefficients preferred (e.g., AC(3,1), AC(1,3)), avoiding low frequencies (visual impact) and high frequencies (easily discarded by compression)
- Strength parameter: Controls embedding energy. 20-40 for daily use; 50+ for extreme compression resistance (social media re-compression)
- Difference amplification: Pixel differences amplified 20× to visualize watermark energy concentrated in mid-frequency textures
Common Use Cases
- Copyright protection: Embed author name or copyright notice in photographs and designs
- Traceability: Embed unique IDs in copies distributed to different recipients to trace leaks
- Anti-fraud verification: Embed verification data in ID photos and receipt images
- Data provenance: Embed markers in AI training datasets to detect unauthorized use
All computation runs locally in your browser. Images never leave your device. Output is lossless PNG to maximally preserve watermark data.