In this blog post, we will learn how watermarks inserted into digital photos are hidden and resistant to editing.
Digital watermarking is the technology of invisibly inserting a specific identifier, or watermark, into digital photos. These watermarks can be extracted using certain methods after insertion and can be used as a means of proving the copyright of a photo.
Therefore, watermarking technology must be able to extract watermarks reliably in a form close to the original, even after the original image has been rotated, cropped, reduced, or otherwise edited, without damaging the watermark. This is called robustness. In other words, durability is necessary to prevent the watermark from being damaged by external factors. In addition, even if a watermark is inserted, the original data storage format must remain unchanged, and invisibility must also be ensured so that the inserted identifier is not easily visible to the user. This is one of the important requirements of watermarking technology.
Digital photo data is generally expressed as brightness values of pixels arranged in a grid pattern in horizontal and vertical directions. The method of arranging the brightness values of each pixel in a two-dimensional array is called the spatial domain method. In this spatial domain method, watermarks can be inserted by finely adjusting the brightness values of pixels in areas that are not easily detected by the human eye.
For example, it is possible to insert image data of a specific brand by adding or multiplying the pixel values at specific locations in the spatial domain. The advantage of the spatial domain method is that the amount of computation is relatively small and the insertion speed of the identifier is fast because the watermark can be inserted directly into the pixel values. In other words, it has excellent processing efficiency.
However, since watermarks are inserted in a specific location, there is a high possibility that they will be damaged or disappear during simple image editing techniques such as cropping or rotation, or during lossy compression processes such as JPEG.
To overcome the disadvantages of the spatial domain method, the frequency domain method is also used. This method is based on the concept of interpreting the degree of spatial brightness change in digital images as spatial frequency. Spatial frequency refers to the degree of vibration in brightness values according to spatial position changes, rather than changes over time. In other words, the more frequently contrast changes occur in a specific direction within a digital image, the higher the spatial frequency measured in that direction. When there is a sudden change in brightness between adjacent pixels, the spatial frequency of that section reaches its maximum value.
Using this principle, digital photos can be expressed as a distribution of spatial frequencies on a two-dimensional plane in the horizontal and vertical directions. The overall distribution of frequency information composed of a two-dimensional array as described above is called the spatial frequency spectrum. This frequency spectrum can be obtained through a mathematical conversion process, and generally, conversion techniques such as Fourier transform are used to convert data from the spatial domain to the frequency domain. This process is lossless and can also be reversed.
In order to insert a watermark into the frequency domain, the image in the spatial domain is first converted into the frequency domain using a mathematical method such as Fourier transform, and then the watermark data is inserted into a specific frequency band. After that, the image is restored to the spatial domain through a reverse conversion process. The key point of this method is that when a watermark is inserted into a specific frequency band, the information is distributed and stored across all pixels that make up that frequency. As a result, the watermark is spread across the entire image, so even if some editing is performed, such as simple cutting or rotation, the watermark can be restored to a certain extent using only the information remaining in the remaining areas.
However, this frequency domain-based method has the disadvantage of requiring much more computation to insert and extract watermarks than the spatial domain method because the conversion process is essential. In addition, watermarks inserted into the frequency domain appear as noise when converted back to the spatial domain, which can cause visual quality degradation across the entire original image. In other words, the image may become blurry or slightly distorted.
Most of the information that humans perceive visually in digital photos is concentrated in the low-frequency band. Humans are more sensitive to low-frequency components than high-frequency components when recognizing the contents of images. In other words, the less contrast there is and the smoother the image, the more it attracts visual attention. On the other hand, the high-frequency band is responsible for contours and detailed textures, and watermarks inserted in this band are difficult to see visually.
Therefore, even if the amount of noise caused by watermarks is the same regardless of the band, watermarks inserted in the high-frequency band are less visually noticeable, making them advantageous for ensuring cognitive invisibility. However, another problem arises. Most lossy compression technologies adopt a method of removing relatively less important high-frequency components to reduce the overall data size of images. JPEG is a typical example of a lossy compression format. As a result, watermarks inserted in the high-frequency band are easily damaged or disappear during the compression process.
For this reason, frequency domain-based watermarking technology often utilizes the mid-frequency band. The mid-frequency band is not easily visible to the human eye and is relatively less susceptible to damage from lossy compression. In other words, it is the band that best maintains the balance between invisibility and robustness. Although the frequency domain method has the disadvantages of high technical complexity and large computation volume, it is widely studied because it offers much better durability and security than the spatial domain method.
