Abstract
Th is innovation is to use Recurrent Neural Network (RNN), a typical machine learning model such as Long Short Term Memory (LSTM) to predict video frames. The RNN model can be trained offline and refined in an online and dynamic fashion. The hidden states in the model can be defined as coding parameters such as frame or macroblock coding model, motion vectors, etc. If a higher frame rate is needed for a smooth video playback, this model can predict and insert frames. If one or more frames are missing due to various reasons, this model can place the predicted frames in the correct places. Furthermore, if the decoded frames are not IDR or I-frame or don't have large number of coded Intra blocks, the predicted frames can be used to verify and report possible quality degradation.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Anonymous, "Frame Recovery by Dynamic Learning and Prediction in a Video Decoder", Technical Disclosure Commons, (April 14, 2021)
https://www.tdcommons.org/dpubs_series/4219