Techniques for processing videos using temporally-consistent transformer model
Techniques are disclosed for enhancing videos using a machine learning model that is a temporally-consistent transformer model. The machine learning model processes blocks of frames of a video in which the temporally first input video frame of each block of frames is a temporally second to last output video frame of a previous block of frames. After the machine learning model is trained, blocks of video frames, or features extracted from the video frames, can be warped using an optical flow technique and transformed using a wavelet transform technique. The transformed video frames are concatenated along a channel dimension and input into the machine learning model that generates corresponding processed video frames.
1 . A computer-implemented method for enhancing videos, the method comprising:
receiving, from a video, a first plurality of video frames and a second plurality of video frames;
processing the first plurality of video frames using a machine learning model to generate a first plurality of processed video frames;
replacing a temporally first video frame included in the second plurality of video frames with a temporally second to last video frame included in the first plurality of processed video frames; and
processing the second plurality of video frames using the machine learning model to generate a second plurality of processed video frames.
2 . The computer-implemented method of claim 1 , further comprising performing one or more operations to train the machine learning model using a loss function that penalizes a difference between a temporally last frame of a plurality of processed training video frames generated by the machine learning model and a temporally first frame of a subsequent plurality of processed training video frames generated by the machine learning model.
3 . The computer-implemented method of claim 1 , wherein processing the first plurality of video frames using the machine learning model comprises:
performing one or more transform operations on each video frame included in the first plurality of video frames to generate a plurality of transformed video frames;
concatenating the plurality of transformed video frames along a channel dimension to generate a concatenated set of transformed video frames; and
inputting the concatenated set of transformed video frames into the machine learning model.
4 . The computer-implemented method of claim 3 , wherein the one or more transform operations include at least one of (1) one or more wavelet transform operations or (2) one or more pixel shuffle operations.
5 . The computer-implemented method of claim 1 , further comprising:
generating an optical flow based on the first plurality of video frames; and
warping either the first plurality of video frames or features extracted from the first plurality of video frames based on the optical flow.
6 . The computer-implemented method of claim 1 , further comprising:
adding a plurality of amounts of degradation to a set of video frames to generate a plurality of sets of degraded video frames, wherein each set of degraded video frames included in the plurality of sets of degraded video frames includes a different amount of the degradation; and
training the machine learning model based on the plurality of sets of degraded video frames.
7 . The computer-implemented method of claim 6 , wherein the degradation comprises at least one of noise or blur.
8 . The computer-implemented method of claim 1 , wherein the machine learning model comprises a transformer model.
9 . The computer-implemented method of claim 1 , wherein the machine learning model comprises one or more three-dimensional (3D) convolution layers.
10 . One or more non-transitory computer-readable storage media including instructions that, when executed by one or more processing units, cause the one or more processing units to perform steps for enhancing videos, the steps comprising:
receiving, from a video, a first plurality of video frames and a second plurality of video frames;
processing the first plurality of video frames using a machine learning model to generate a first plurality of processed video frames;
replacing a temporally first video frame included in the second plurality of video frames with a temporally second to last video frame included in the first plurality of processed video frames; and
processing the second plurality of video frames using the machine learning model to generate a second plurality of processed video frames.
11 . The one or more non-transitory computer-readable storage media of claim 10 , wherein the instructions, when executed by the one or more processing units, further cause the one or more processing units to perform the step of:
performing one or more operations to train the machine learning model using a loss function that penalizes a difference between a temporally last frame of a plurality of processed training video frames generated by the machine learning model and a temporally first frame of a subsequent plurality of processed training video frames generated by the machine learning model.
12 . The one or more non-transitory computer-readable storage media of claim 10 , wherein processing the first plurality of video frames using the machine learning model comprises:
performing one or more transform operations on each video frame included in the first plurality of video frames to generate a plurality of transformed video frames;
concatenating the plurality of transformed video frames along a channel dimension to generate a concatenated set of transformed video frames; and
inputting the concatenated set of transformed video frames into the machine learning model.
13 . The one or more non-transitory computer-readable storage media of claim 12 , wherein the one or more transform operations include at least one of one or more wavelet transform operations or one or more pixel shuffle operations.
14 . The one or more non-transitory computer-readable storage media of claim 10 , wherein the instructions, when executed by the one or more processing units, further cause the one or more processing units to perform the steps of:
generating an optical flow based on the first plurality of video frames; and
warping either the first plurality of video frames or features extracted from the first plurality of video frames based on the optical flow.
15 . The one or more non-transitory computer-readable storage media of claim 10 , wherein the instructions, when executed by the one or more processing units, further cause the one or more processing units to perform the steps of:
adding a plurality of amounts of degradation to a set of video frames to generate a plurality of sets of degraded video frames, wherein each set of degraded video frames included in the plurality of sets of degraded video frames includes a different amount of the degradation; and
training the machine learning model based on the plurality of sets of degraded video frames.
16 . The one or more non-transitory computer-readable storage media of claim 15 , wherein the degradation comprises at least one of noise or blur.
17 . The one or more non-transitory computer-readable storage media of claim 10 , wherein the machine learning model comprises a transformer model.
18 . The one or more non-transitory computer-readable storage media of claim 10 , wherein the machine learning model comprises one or more three-dimensional (3D) convolution layers.