Video Codec Using Machine Learning Image Compression Techniques

The paper proposes a video codec using an artificial intelligence-based image compression technique. The video codec generalizes an image encoding technique that uses the redefinition of the texture parts of an image. The video codec uses GAN-based artificial intelligence algorithms to create synthetic texture parts of a video frame. In addition, the video codec uses dispersed views of the video frame to encode meaningful parts of the video frame (outside the texture parts). The dispersed view of the video frame includes several dispersed tiers. Each next tier is twice as detailed as the previous one. The video codec encodes the dispersed tiers of the video frame one by one, from the most dispersed tier to the least dispersed tier (within (meaningful parts of the video frame). The video codec approximates each dispersed tier of the video frame based on the more dispersed tiers through artificial intelligence (the video codec may use a super-resolution neural network or a decision tree). The video codec then archives the approximation errors into an encoded data file. In addition, the proposed video codec moves from a two-dimensional to a three-dimensional approximator when generalizing the image-encoding algorithm to the case of video encoding. The three-dimensional approximator allows you to exploit the interdependence of video frames. In addition, the proposed video codec also uses artificial intelligence to predict video frames. Computer simulations in natural videos confirm noticeable gains of the proposed algorithm (up to 26% in terms of the encoded data size).