Method and apparatus of neural network for video coding
Method and apparatus of video encoding video coding for a video encoder or decoder using Neural Network (NN) are disclosed. According to one method, input data or a video bitstream are received for blocks in one or more pictures, which comprise one or more colour components. The residual data, prediction data, reconstructed data, filtered-reconstructed data or a combination thereof is derived for one or more blocks of said one or more pictures. A target signal corresponding to one or more of the about signal types is processed using a NN (Neural Network) and the input of the NN or an output of the NN comprises two or more colour components. According to another method, A target signal corresponding to one or more of the about signal types is processed using a NN and the input of the NN or an output of the NN comprises two or more colour components.
1. A method of video coding for a video encoder or decoder, the method comprising:
receiving input data at an encoder side or receiving a video bitstream at a decoder side, wherein the input data correspond to a picture in a video sequence or the video bitstream comprises compressed data associated with the picture in the video sequence;
deriving residual data and prediction data for a block of the picture based on the input data or the compressed data;
generating reconstructed data for the block by combining the residual data and the prediction data for the block;
performing a NN (Neural Network) process to generate filtered reconstructed data for the block according to inputs of the NN that include, provided as separate input channels of the inputs:
(i) the reconstructed data for the block, and
(ii) the residual data or the prediction data for the block; and
generating reconstructed picture data of the picture based on the filtered reconstructed data for the block.
2. The method of claim 1 , wherein the NN process is configured to generate two chroma components of the filtered reconstructed data for the block.
3. The method of claim 1 , wherein
the NN process is configured to generate a luma component of the filtered reconstructed data for the block according to the inputs of the NN that include a luma component of the reconstructed data for the block and a chroma component of the reconstructed data for the block, and
the chroma component of the reconstructed data for the block is up-sampled to match a sampling rate of the luma component of the reconstructed data for the block.
4. The method of claim 1 , wherein the NN process is configured to generate one luma component of the filtered reconstructed data for the block and two chroma components of the filtered reconstructed data for the block.
5. The method of claim 1 , wherein
the NN process is configured to generate a chroma component of the filtered reconstructed data for the block according to the inputs of the NN that include a chroma component of the reconstructed data for the block and a luma component of the reconstructed data for the block, and
the luma component of the reconstructed data for the block is down-sampled to match a sampling rate of the chroma component of the reconstructed data for the block.
6. The method of claim 1 , wherein the NN corresponds to CNN (convolutional neural network), DNN (deep fully-connected feed-forward neural network), RNN (recurrent neural network), or GAN (generative adversarial network).
7. The method of claim 1 , wherein
the NN process is configured to generate two or more color components of the filtered reconstructed data for the block, and
the method further comprises determining a shared control flag indicative of enabling the NN process for the two or more color components of the filtered reconstructed data for the block.
8. An apparatus for video coding in a video encoder or decoder, the apparatus comprising one or more electronics or processors configured to:
receive input data at an encoder side or receive a video bitstream at a decoder side, wherein the input data correspond to a picture in a video sequence or the video bitstream comprises compressed data associated with the picture in the video sequence;
derive residual data and prediction data for a block of the picture based on the input data or the compressed data;
generate reconstructed data for the block by combining the residual data and the prediction data for the block;
perform a NN (Neural Network) process to generate filtered reconstructed data for the block according to inputs of the NN that include, provided as separate input channels of the inputs:
(i) the reconstructed data for the block, and
(ii) the residual data or the prediction data for the block; and
generate reconstructed picture data of the picture based on the filtered reconstructed data for the block.
9. A method of video coding for a video encoder or decoder, the method comprising:
receiving input data at an encoder side or receiving a video bitstream at a decoder side, wherein the input data correspond to a picture in a video sequence or the video bitstream comprises compressed data associated with the picture in the video sequence;
deriving residual data and prediction data for a block of the picture based on the input data or the compressed data;
generating reconstructed data for the block by combining the residual data and the prediction data for the block;
deriving pieces of data corresponding to respective portions of the block from a color component of the reconstructed data for the block;
performing a NN (Neural Network) process to generate filtered reconstructed data for the block according to inputs of the NN that include at least the pieces of data as separate input channels of the inputs of the NN; and
generating reconstructed picture data of the picture based on the filtered reconstructed data for the block.
10. The method of claim 9 , wherein the portions of the block correspond to subsampling the block according to different phases.
11. The method of claim 9 , wherein the NN corresponds to CNN (convolutional neural network), DNN (deep fully-connected feed-forward neural network), RNN (recurrent neural network), or GAN (generative adversarial network).
12. The method of claim 9 , wherein
the NN process is configured to generate a luma component of the filtered reconstructed data for the block according to the inputs of the NN that include a luma component of the reconstructed data for the block and a chroma component of the reconstructed data for the block, and
the chroma component of the reconstructed data for the block is up-sampled to match a sampling rate of the luma component of the reconstructed data for the block.
13. The method of claim 9 , wherein
the color component of the reconstructed data for the block corresponds to a luma component of the reconstructed data for the block, and
the NN process is configured to output a luma component of the filtered reconstructed data for the block in data pieces that correspond to the respective portions of the block.
14. The method of claim 9 , wherein the NN process is configured to output more than one color components of the filtered reconstructed data for the block,
the color component of the reconstructed data for the block corresponds to a luma component of the reconstructed data for the block,
the inputs of the NN further include a chroma component of the reconstructed data for the block as another input channel of the inputs of the NN, and
the NN process is configured to output (i) a chroma component of the filtered reconstructed data for the block and (ii) a luma component of the filtered reconstructed data for the block in data pieces that correspond to the respective portions of the block.
15. The method of claim 9 , wherein
the NN process is configured to generate a chroma component of the filtered reconstructed data for the block according to the inputs of the NN that include a chroma component of the reconstructed data for the block and a luma component of the reconstructed data for the block, and
the luma component of the reconstructed data for the block is down-sampled to match a sampling rate of the chroma component of the reconstructed data for the block.
16. The method of claim 9 , wherein the inputs of the NN further include the residual data or the prediction data for the block.
17. An apparatus for video coding in a video encoder or decoder, the apparatus comprising one or more electronics or processors configured to:
receive input data at an encoder side or receiving a video bitstream at a decoder side, wherein the input data correspond to a picture in a video sequence or the video bitstream comprises compressed data associated with the picture in the video sequence;
derive residual data and prediction data for a block of the picture based on the input data or the compressed data;
generate reconstructed data for the block by combining the residual data and the prediction data for the block;
derive pieces of data corresponding to respective portions of the block from a color component of the reconstructed data for the block;
perform a NN (Neural Network) process to generate filtered reconstructed data for the block according to inputs of the NN that include at least the pieces of data as separate input channels of the inputs of the NN; and
generate reconstructed picture data of the picture based on the filtered reconstructed data for the block.
18. The apparatus of claim 17 , wherein the inputs of the NN further include the residual data or the prediction data for the block.