CamCyclopedia
Contents
- Summary
- Use
- How to create GIFs?
- Quantization
- Dithering
- GIF remaster details
Summary
Stephen Walhite first proposed the idea of GIFs in 1987. It quickly gained widespread popularity with the advent of the Internet, especially with the 1989 revised GIF89a adding animation capabilities.
GIFs have unique advantages, such as the ability to display transparent backgrounds, and are still in active use today. But, as an older format, they have quality issues. GIF Remaster applies the latest deep learning techniques to improve the quality of GIF images.
Use
GIF Remaster works with One UI 5.1 and above, and first debuted on the Galaxy S23 series. We are working on expanding support to more models. It works just like any other Remaster feature. You can use it by accessing Gallery suggestions or the Photo Viewer.
How to create GIFs?
GIF images are created by passing the original image through two main stages: lossy compression and lossless compression. The second stage, lossless compression, uses LZW coding, which is a type of entropy coding, similar to ZIP file compression, and results in no information loss. The first stage, lossy compression, results in information loss and degradation.
Quantization
GIFs remap the input pixel colors into a smaller color palette of below 8-bit 256-size. This is called quantization. A JPEG image depicts 16,777,216 colors at 24 bits per pixel, and the marked differences from the original, when converting it to a GIF, are quite noticeable and frustrating.
Dithering
One way to hide the unpleasantness caused by quantization errors is to add artificial noise, known as dithering. Dithering can smooth out the boundaries between quantized regions, but adding noise by the book may also introduce unwanted noise.
GIF Remaster details
GIF Remaster utilizes deep learning technology to improve quality. Deep learning training is very effective at improving GIFs with different types and levels of quality degradation by referencing large amounts of data.
Noise reduction and upscaling
The noise generated by quantization and dithering is highly variable depending on the parameters set during GIF creation. The magnitude of the noise is quite large compared to later encoding methods. GIF remastering with deep learning effectively attenuates both types of noise. In addition, GIF images usually have a low resolution. GIF Remaster up-scales them by up to 2x to add more detail and sharpness.
Transparent background GIF pre-processing
Passing a transparent background image straight through a deep learning network will result in unwanted noise on the pixels at the edge of an object, as shown in the image below (without pre-processing). This is because the deep learning network also references the colors of the adjacent pixels when creating the edge pixels of an object. Since the transparent background pixels have an opacity of 0 to make them invisible, but their RGB value information (0, which is mostly black) is still valid, the RGB information of these transparent pixels influences the deep learning network to create noise. GIF Remaster applies an algorithm to process the transparent background pixels so that the deep learning network is not affected by them, resulting in an output with no boundary noise as shown in the [pre-processing applied] image below. This technology is currently patented.
■ Related Content
Remaster Picture - Samsung Members
CamCyclopedia Index - Samsung Members
You can also check out CamCyclopedia at any time by going to Community -> Category (app) -> CamCyclopedia -> "CamCyclopedia Index".
2 Comments
