Convolutional-neutral-network based filter for video coding
The present disclosure provides methods for convolutional-neural-network (CNN) based filter for video coding. An exemplary method includes: applying motion estimation to a target coding block, to determine a reference block of the target coding block; inputting, to a convolutional neural network (CNN) filter, image data associated with the target coding block and the reference block; and executing the CNN filter to determine a residual associated with the target coding block based on the input image data.
1. A computer-implemented video encoding method, comprising:
determining a first rate distortion associated with a target coding block based on a convolutional neural network (CNN) filter being disabled, the CNN filter being configured to determine a residual associated with the target coding block based on image data associated with the target coding block and a reference block of the target coding block, the reference block being determined by applying motion estimation to the target coding block;
determining a second rate distortion associated with the target coding block based on the CNN filter being enabled;
perform a comparison of the first rate distortion and the second rate distortion; and
based on the comparison, encoding the target coding block and setting a flag indicating whether the CNN filter is to be enabled or disabled.
2. The method according to claim 1 , wherein the flag is signaled in at least one of a picture parameter set (PPS), a picture header, or a slice header.
3. The method according to claim 1 , wherein the flag is a frame-level flag, and when the frame-level flag indicates that the CNN filter is to be disabled, the CNN filter is disabled for a frame associated with the frame-level flag.
4. The method according to claim 1 , wherein the flag is a CTU-level flag, and when the CTU-level flag indicates that the CNN filter is to be disabled, the CNN filter is disabled for a CTU associated with the CTU-level flag.
5. The method according to claim 4 , wherein the flag is signaled in a table that comprises a plurality of CTU-level flags corresponding to a plurality of CTUs respectively, each of the plurality of CTU-level flags indicating whether the CNN filter is to be enabled or disabled for the corresponding CTU.
6. The method according to claim 1 , wherein the CNN filter comprises a plurality of convolutional layers, each of the plurality of convolutional layers using one or more filter kernels to generate one or more feature maps respectively, and executing the CNN filter to determine the residual associated with the target coding block based on input image data comprises:
extracting, at a first convolutional layer of the CNN filter, spatial features of the image data associated with the target coding block and the reference block.
7. The method according to claim 1 , wherein the CNN filter is trained using a training data set comprising:
a motion-compensated training block, and
an un-motion-compensated extended region of the motion compensated training block,
the motion-compensated training block being a predicted block based on a motion vector and a reference picture, and the un-motion-compensated extended region comprising unpredicted image data that is not predicted based on the motion vector or reference picture.
8. A computer-implemented video decoding method, comprising:
decoding a video bitstream to generate a reconstructed coding block;
applying motion estimation to the reconstructed coding block, to determine a reference block of the reconstructed coding block;
determining, based on a flag, whether a convolutional neural network (CNN) filter is enabled or disabled; and
executing the CNN filter to determine, based on the reference block, a residual associated with the reconstructed coding block when the CNN filter is enabled,
wherein the flag is generated based on a comparison of a first rate distortion and a second rate distortion associated with the coding block,
wherein the first rate distortion is determined based on the CNN filter being disabled, and
wherein the second rate distortion is determined based on the CNN filter being enabled.
9. The method according to claim 8 , wherein the flag is signaled in at least one of a picture parameter set (PPS), a picture header, or a slice header.
10. The method according to claim 8 , wherein the flag is a frame-level flag, and the method further comprises: in response to the frame-level flag indicating that the CNN filter is disabled, determining that the CNN filter is disabled for a frame associated with the frame-level flag.
11. The method according to claim 8 , wherein the flag is a CTU-level flag, and the method further comprises: in response to the CTU-level flag indicating that the CNN filter is disabled, determining that the CNN filter is disabled for a CTU associated with the CTU-level flag.
12. The method according to claim 11 , wherein the flag is signaled in a table that comprises a plurality of CTU-level flags corresponding to a plurality of CTUs respectively, each of the plurality of CTU-level flags indicating whether the CNN filter is enabled or disabled for the corresponding CTU.
13. The method according to claim 8 , wherein the CNN filter comprises a plurality of convolutional layers, each of the plurality of convolutional layers using one or more filter kernels to generate one or more feature maps respectively, and executing the CNN filter to determine, based on the reference block, a residual associated with the reconstructed coding block comprises:
extracting, at a first convolutional layer of the CNN filter, spatial features of the reconstructed coding block and the reference block.
14. The method according to claim 8 , wherein the CNN filter is trained using a training data set comprising:
a motion-compensated training block, and
an un-motion-compensated extended region of the motion compensated training block,
the motion-compensated training block being a predicted block based on a motion vector and a reference picture, and the un-motion-compensated extended region comprising unpredicted image data that is not predicted based on the motion vector or reference picture.
15. A computer-implemented method of storing a bitstream of a video, the method comprising:
determining a first rate distortion associated with a target coding block based on a convolutional neural network (CNN) filter being disabled, the CNN filter being configured to determine a residual associated with the target coding block based on image data associated with the target coding block and a reference block of the target coding block, the reference block being determined by applying motion estimation to the target coding block;
determining a second rate distortion associated with the target coding block based on the CNN filter being enabled;
perform a comparison of the first rate distortion and the second rate distortion;
generating, based on the comparison, a bitstream comprising a flag indicating whether the CNN filter is to be enabled or disabled; and
storing the bitstream in a non-transitory computer readable medium.
16. The method according to claim 15 , wherein the flag is set in at least one of a picture parameter set (PPS), a picture header, or a slice header.
17. The method according to claim 15 , wherein the flag is a frame-level flag, and when the frame-level flag indicates that the CNN filter is to be disabled, the CNN filter is disabled for a frame associated with the frame-level flag.
18. The method according to claim 15 , wherein the flag is a CTU-level flag, and when the CTU-level flag indicates that the CNN filter is to be disabled, the CNN filter is disabled for a CTU associated with the CTU-level flag.
19. The method according to claim 18 , wherein the flag is set in a table that comprises a plurality of CTU-level flags corresponding to a plurality of CTUs respectively, each of the plurality of CTU-level flags indicating whether the CNN filter is to be enabled or disabled for the corresponding CTU.
20. The method according to claim 15 , wherein the CNN filter comprises a plurality of convolutional layers, each of the plurality of convolutional layers using one or more filter kernels to generate one or more feature maps respectively, and executing the CNN filter to determine the residual associated with the target coding block based on input image data comprises:
extracting, at a first convolutional layer of the CNN filter, spatial features of the image data associated with the target coding block and the reference block.