Deinterlacing via deep learning

1. A method for performing deinterlacing, comprising:

separating a first interlaced video frame into a first sequence of fields ordered by time, the first sequence of fields including a first field representing a first point in time and a second field representing a second point in time that is different from the first point in time;

generating, by applying a machine learning model comprising a deinterlacing network that includes one or more residual links to the first field in the first sequence of fields, a third field that is missing from the first sequence of fields and is complementary to the first field; and

constructing a progressive video frame based on the first field and the third field.

2. The method of claim 1 , further comprising:

extracting a second sequence of fields from a progressive video;

inputting the second sequence of fields as training data for the deinterlacing network; and

updating parameters of the deinterlacing network to reduce an error between a third sequence of fields generated by the deinterlacing network based on the second sequence of fields and a fourth sequence of fields that is complementary to the second sequence of fields in the progressive video.

3. The method of claim 1 , further comprising:

adding, to the first sequence of fields, one or more fields from one or more interlaced video frames adjacent to the first interlaced video frame in an interlaced video; and

applying the deinterlacing network to the first sequence of fields to generate the third field.

4. The method of claim 3 , wherein applying the deinterlacing network to the first sequence of fields comprises:

applying a feature extraction layer to the first sequence of fields to generate a combined feature map; and

inputting the combined feature map into the deinterlacing network.

5. The method of claim 3 , wherein applying the deinterlacing network to the first sequence of fields comprises:

applying a flow estimation model to the first sequence of fields to generate one or more motion vectors between pixels from consecutive fields in the first sequence of fields; and

inputting the one or more motion vectors with the first sequence of fields into the deinterlacing network.

6. The method of claim 3 , wherein the one or more fields comprise a fourth field that precedes or succeeds the first field within the first sequence of fields.

7. The method of claim 1 , further comprising applying an attention mechanism to the third field and additional estimates of the third field to generate an update to the third field prior to constructing the progressive video frame.

8. The method of claim 7 , wherein the attention mechanism comprises a first convolutional branch that extracts features from the third field and the additional estimates of the third field, and a second convolutional branch that calculates attention weights for scaling the features.

9. The method of claim 1 , wherein the deinterlacing network comprises a sequence of dense compression units, each of the dense compression units comprising a set of densely connected layers and a convolutional layer.

10. The method of claim 1 , wherein the one or more residual links include a first residual link from a first input to the deinterlacing network to a first output of the deinterlacing network, and a second residual link from a second input of a dense compression unit to a second output of the dense compression unit.

11. The method of claim 1 , wherein applying the deinterlacing network to the first field comprises inputting a vertical flip of the first field into the deinterlacing network.

12. The method of claim 1 , wherein constructing the progressive video frame comprises combining the first field and the third field into the progressive video frame.

13. A non-transitory computer readable medium storing instructions that, when executed by a processor, cause the processor to perform the steps of:

separating a first interlaced video frame into a first sequence of fields ordered by time, the first sequence of fields including a first field representing a first point in time and a second field representing a second point in time that is different from the first point in time;

adding, to the first sequence of fields, one or more fields from one or more frames adjacent to the first interlaced video frame within an interlaced video;

generating, by applying a machine learning model that comprises a deinterlacing network to the first sequence of fields, a third field that is missing from the first sequence of fields and is complementary to the first field in a progressive video frame; and

constructing a progressive video frame based on the first field and the third field.

14. The non-transitory computer readable medium of claim 13 , wherein the steps further comprise:

extracting a second sequence of fields from a progressive video;

inputting the second sequence of fields as training data for the deinterlacing network; and

updating parameters of the deinterlacing network to reduce an error between a third sequence of fields generated by the deinterlacing network from the second sequence of fields and a fourth sequence of fields that is complementary to the second sequence of fields in the progressive video.

15. The non-transitory computer readable medium of claim 13 , wherein

the first sequence of fields comprises a fourth field that precedes or succeeds the first field within the interlaced video.

16. The non-transitory computer readable medium of claim 13 , wherein applying the deinterlacing network to the first sequence of fields comprises:

applying a feature extraction layer to the first sequence of fields to generate a combined feature map; and

inputting the combined feature map into the deinterlacing network.

17. The non-transitory computer readable medium of claim 13 , wherein applying the deinterlacing network to the first sequence of fields comprises:

applying a flow estimation model to the first sequence of fields to generate one or more motion vectors between pixels from consecutive fields in the first sequence of fields;

generating a motion compensated estimation of the third field based on the one or more motion vectors; and

inputting the motion compensated estimation of the third field with the first sequence of fields into the deinterlacing network.

18. The non-transitory computer readable medium of claim 13 , wherein the steps further comprise applying an attention mechanism to the third field and additional estimates of the third field to generate an update to the third field prior to constructing the progressive video frame.

19. The non-transitory computer readable medium of claim 13 , wherein constructing the progressive video frame comprises obtaining the progressive video frame as output from the deinterlacing network.

20. A system, comprising:

a memory that stores instructions, and

a processor that is coupled to the memory and, when executing the instructions, is configured to:

separate a first interlaced video frame into a first sequence of fields ordered by time, the first sequence of fields including a first field representing a first point in time and a second field representing a second point in time that is different from the first point in time;

generate, by applying a machine learning model comprising a deinterlacing network that includes one or more residual links to the first field in the first sequence of fields, a third field that is missing from the first sequence of fields and is complementary to the first field; and

construct a progressive video frame based on the first field and the third field.