System and method for multi-frame contextual attention for multi-frame image and video processing using deep neural networks

1. A method for multi-frame contextual attention, the method comprising:

obtaining a reference frame from a video, the reference frame having an occluded region;

identifying context frames from the video with respect to the reference frame, the context frames comprising a first set of one or more consecutive neighboring frames immediately prior to the reference frame and a second set of one or more consecutive neighboring frames immediately subsequent to the reference frame, wherein a first number of frames in the first set is equal to a second number of frames in the second set, and the first number and the second number have an inversely proportional relationship to a scene changing speed in the video;

producing a coarse prediction of the reference frame using at least the reference frame to fill a hole of the occluded region based on filtering; and

producing a refined reference frame by processing the coarse prediction of the reference frame and the context frames.

2. The method of claim 1 , wherein producing the refined reference frame further includes producing a coarse prediction of each context frame based on convolutional filtering.

3. The method of claim 2 , wherein producing the refined reference frame further includes processing the coarse predictions of each context frame and the coarse prediction of the reference frame in a multi-frame contextual attention stream.

4. The method of claim 1 , wherein producing the refined reference frame further includes processing the coarse prediction of the reference frame and the identified context frames in a multi-frame contextual attention stream.

5. The method of claim 1 , wherein the identified context frames include non-masked regions in feature maps corresponding to multiple frames.

6. The method of claim 1 , further comprising detecting the amount of motion in the video.

7. The method of claim 1 , wherein the reference frame is occluded by an object, and further comprising:

removing the object producing the hole in the reference frame,

wherein processing the coarse prediction of the reference frame and the context frames includes filling the hole in the obtained reference frame.

8. A system for multi-frame contextual attention, the system comprising:

a memory; and

a processor configured to:

obtain a reference frame from a video, the reference frame having an occluded region;

identify context frames from the video with respect to the reference frame, the context frames comprising a first set of one or more consecutive neighboring frames immediately prior to the reference frame and a second set of one or more consecutive neighboring frames immediately subsequent to the reference frame, wherein a first number of frames in the first set is equal to a second number of frames in the second set, and the first number and the second number have an inversely proportional relationship to a scene changing speed in the video; and

produce a coarse prediction of the reference frame using at least the reference frame to fill a hole of the occluded region through filtering;

produce a refined reference frame by processing the coarse prediction of the reference frame and the context frames.

9. The system of claim 8 , wherein producing the refined reference frame further includes producing a coarse prediction of each context frame based on convolutional filtering.

10. The system of claim 9 , wherein the processor is further configured to produce the refined reference frame by processing the coarse predictions of each context frame and the coarse prediction of the reference frame in a multi-frame contextual attention stream.

11. The system of claim 8 , wherein the processor is further configured to produce the refined reference frame by processing the coarse prediction of the reference frame and the identified context frames in a multi-frame contextual attention stream.

12. The system of claim 8 , wherein the identified context frames include non-masked regions in feature maps corresponding to multiple frames.

13. The system of claim 8 , wherein the processor is further configured to detect the amount of motion in the video.

14. The system of claim 8 , wherein the reference frame is occluded by an object, and

wherein the processor is further configured to remove the object producing a hole in the reference frame, and processing the coarse prediction of the reference frame and the context frames includes filling the hole in the obtained reference frame.