How to achieve high-definition super-resolution algorithm to achieve 480P 4K
I believe that many people like watching movies, animation, etc., whether they are watching or watching, high-definition, high-resolution videos in local viewing, and bring a better viewing experience. But for those nostalgia, it may be not so lucky. In the past, the resolution of many old movies, old movies, and old movies may not reach 720P, which leads to the experience of watching experience, after all, in this age, low resource, low resource It is really a bit out of date.
In fact, even if you look at the entire industry, the lack of super high-definition content is also the pain point where the industry is generally existed. The incomplete shooting, shooting, production level is immature, the shortage of technology has become a stumbling block that hinders the development of the industry, in order to improve ultra HD The production capacity of the video, while saving costs, requiring artificial intelligence intervention, super-resolution algorithm is a good way to solve this problem.
It is not familiar with the super-resolution algorithm. Many people in front of the screen may not be familiar, but one mentioned DLSS, FSR or XESS, the game players must have earned, although they are not the same, but from the results In view, these three technologies can improve game resolution and bring better gaming experience. The super-resolution algorithm to be said today, in a sense, and the AMD's FSR technology can be said to be quite similar.
Image super resolution problems Research is how to get a high-resolution image when entering a low-resolution image, and the traditional image interpolation algorithm can give this effect to some extent, such as nearest neighbor interpolation, double line Sex interpolation and double interpolation, etc., but these algorithms achieved high resolution image effects are not ideal.
In terms of image processing, another famous algorithm WAIFU2X, which uses the SRCNN convolutional algorithm, SRCNN is the first super-resolution algorithm using a CNN structure (ie, based on deep learning), it will use the depth of the algorithm process The method of learning is achieved, and the effect is better than the traditional multi-module integration. The SRCNN process is as follows: First enter the pre-processing. The entered low resolution LR image is enlarged using a Bicubic algorithm to enlarge the target size. Then the target of the next algorithm is to obtain an image of the input comparative blurred LR, through the process of convolution network, to obtain an image of the super-resolution SR, so that it is similar to the high-resolution HR image of the original image.
The AI super resolution technology is a direction in the field of image repair technology. During the output of anime video, there is often a series of digital signal processing, including serrated, halo, color block, noise treatment, blur line processing, etc., in the past, video workers often need to sample the source, in Different fragments are analyzed at the master belt resolution, and the manual repair is made by a series of filters, which causes great human cost.
Today's protagonist is the real-cugan tool for the B station AI laboratory (project address:https://github.com/bilibili/ailab/tree/main/real-cugan), As long as it uses it, the quality of the animation image can be increased by 2 to 4 times, and it is almost ineffective.
Real-cugan's full name REAL CASCADED-U-Net-Style Generative Adversarial Networks (real, cascaded U-NET style generation confrontation network), using the same animation network structure as WAIFU2X, but because of new training Data and training methods, thereby forming different parameters and reasoning methods.
From the technical details, Real-Cugan will process the anime frame first, then use the image quality score model to score the candidate block to obtain a million-level high-quality anime image block training set. Then use a multi-stage degradation algorithm to obtain a hypervarium image to obtain a low-quality image by allowing the AI model to learn, optimize the reconstruction process of low-quality images to high quality images. After training, the real secondary element is low. Image performs high-definition processing.
Real-cugan supports 2x \ 3x \ 4x times super resolution, which supports 4 noise reduction intensity and conservative fixes, 3 times / 4 times model supports 2 noise reduction intensity and conservative fixes, and if you are Windows users, the author is also ready to prepare the Windows-GUI version.https://github.com/justin62628/squirrel-rife/releases/tag/v0.0.3 ,downloadWork can be used later.
Compared with WAIFU2X (CUNET) and Real-Esrgan (Anime6b), the advantage of Real-Cugan is also more obvious. The author has also conducted a wave of comparison, as follows:
Currently the OGV country in B station"Season 2 of Zhen Soul Street"It has already launched a 4K resolution version of the over-depth. I believe that in the future, more HD reset versions will also be on the road, the development of AI technology is being enhanced from all angles.