System and method for enhancing resolution of video content

1. A computer-implemented method for enhancing video resolution, the method comprising:

receiving, by a video processing system, a Low-Resolution (LR) video, the LR video comprising a plurality of LR images; and

converting, by the video processing system, the LR video to a High-Resolution (HR) video by transforming at least one LR image from among the plurality of LR images to a HR image, an LR image from among the at least one LR image transformed to the HR image by performing the steps of:

performing pixel unshuffle to rearrange image portions in the LR image to generate a set of downsampled sub-images corresponding to the LR image;

extracting shallow features from the set of downsampled sub-images, wherein the shallow features extracted corresponding to the set of downsampled sub-images configure a feature space corresponding to the set of downsampled sub-images;

processing the shallow features using a series of inverted residual blocks to generate enhanced features corresponding to the set of downsampled sub-images, wherein the shallow features are gradually modified in the feature space by the series of inverted residual blocks; and

mapping the enhanced features to a high resolution feature space using pixel shuffle to generate the HR image corresponding to the LR image.

2. The method as claimed in claim 1 , wherein the conversion of the LR video to the HR video is performed in substantially real-time in relation to a playback request for streaming content corresponding to the LR video.

3. The method as claimed in claim 2 , further comprising:

facilitating, by the video processing system, display of the HR video in relation to the playback request for streaming content corresponding to the LR video, wherein the HR video is displayed on a display screen of an electronic device associated with a content viewer.

4. The method as claimed in claim 1 , wherein each inverted residual block is configured to perform at least a first convolution and a second convolution to transform the feature space corresponding to the set of downsampled sub-images from a low-dimensional feature space to an enhanced feature space, wherein the enhanced feature space comprises the enhanced features corresponding to the set of downsampled sub-images.

5. The method as claimed in claim 4 , wherein a Rectified Linear Unit (ReLU) activation function is used for computing output of the first convolution and the second convolution.

6. The method as claimed in claim 1 , wherein the shallow features are extracted from the set of downsampled sub-images using a 3×3 convolution layer.

7. The method as claimed in claim 1 , wherein each inverted residual block of the series of inverted residual blocks is configured to expand the shallow features by an expansion factor to generate the enhanced features.

8. The method as claimed in claim 1 , wherein the video processing system comprises a neural network (NN) for facilitating conversion of the LR video to the HR video, the neural network comprising a plurality of layers.

9. The method as claimed in claim 8 , wherein the plurality of layers comprise at least one of a pixel unshuffle layer, one or more convolutional layers, an inverted residual layer, and a pixel shuffle layer.

10. The method as claimed in claim 8 , further comprising:

decomposing, by the video processing system, the HR image into a plurality of image patches; and

determining, by the video processing system, a weight of each image patch from among the plurality of image patches,

wherein a standard deviation of weights assigned to sub-patch portions of each image patch is computed to determine the weight of the respective image patch, and

wherein the weight of each image patch is indicative of a super resolution restoration difficulty associated with the respective image patch.

11. The method as claimed in claim 10 , further comprising:

computing, by the video processing system, a loss function based on the weight assigned to each image patch from among the plurality of image patches, wherein the loss function is formulated as:

L

⁡

(

θ

)

=

1

N

⁢

∑

i

=

0

N

W

i

⁢



f

S

⁢

R

(

I

L

⁢

R

i

)

-

(

I

H

⁢

R

i

)



where θ denotes parameters of the neural network f SR , N is the number of training samples, W i is the computed weight according to i-th image patch, I LR i and I HR i denote the i-th image patch and corresponding ground truth.

12. The method as claimed in claim 11 , further comprising:

optimizing, by the video processing system, the neural network to facilitate conversion of the LR video to the HR video, wherein the neural network is optimized based on the computed loss function.

13. A video processing system for enhancing video resolution, the video processing system comprising:

a memory for storing instructions; and

a processor configured to execute the instructions and thereby cause the video processing system to at least perform:

receiving a Low-Resolution (LR) video, the LR video comprising a plurality of LR images; and

converting the LR video to a High-Resolution (HR) video by transforming at least one LR image from among the plurality of LR images to a HR image, an LR image from among the at least one LR image transformed to the HR image by performing the steps of:

performing pixel unshuffle to rearrange image portions in the LR image to generate a set of downsampled sub-images corresponding to the LR image;

extracting shallow features from the set of downsampled sub-images, wherein the shallow features extracted corresponding to the set of downsampled sub-images configure a feature space corresponding to the set of downsampled sub-images;

processing the shallow features using a series of inverted residual blocks to generate enhanced features corresponding to the set of downsampled sub-images, wherein the shallow features are gradually modified in the feature space by the series of inverted residual blocks; and

mapping the enhanced features to a high resolution feature space using pixel shuffle to generate the HR image corresponding to the LR image.

14. The video processing system as claimed in claim 13 , wherein the conversion of the LR video to the HR video is performed in real-time in relation to a playback request for streaming content corresponding to the LR video.

15. The video processing system as claimed in claim 14 , wherein the video processing system is further caused to:

facilitate display of the HR video in relation to the playback request for the streaming content corresponding to the LR video, wherein the HR video is displayed on a display screen of an electronic device associated with a content viewer.

16. The video processing system as claimed in claim 13 , wherein each inverted residual block is configured to perform at least a first convolution and a second convolution to transform the feature space corresponding to the set of downsampled sub-images from a low-dimensional feature space to an enhanced feature space, wherein the enhanced feature space comprises the enhanced features corresponding to the set of downsampled sub-images.

17. The video processing system as claimed in claim 13 , wherein the video processing system comprises a neural network (NN) for facilitating conversion of the LR video to the HR video, the neural network comprising a plurality of layers.

18. The video processing system as claimed in claim 17 , wherein each inverted residual block is configured to:

decompose the HR image into a plurality of image patches;

determine a weight of each image patch from among the plurality of image patches, wherein a standard deviation of weights assigned to sub-patch portions of each image patch is computed to determine the weight of a respective image patch, and wherein the weight of each image patch is indicative of a super resolution restoration difficulty associated with the respective image patch;

compute a loss function based on the weight assigned to each image patch from among the plurality of image patches; and

optimize the neural network to facilitate conversion of the LR video to the HR video, wherein the neural network is optimized based on the computed loss function.

19. A computer-implemented method for enhancing video resolution, the method comprising:

receiving, by a video processing system, a Low-Resolution (LR) video in response to a playback request for streaming content provided by a content provider, the LR video comprising a plurality of LR images;

converting, by the video processing system, the LR video to a High-Resolution (HR) video in substantially real-time by transforming at least one LR image from among the plurality of LR images to a HR image, an LR image from among the at least one LR image transformed to the HR image based on processing a set of downsampled sub-images configuring the LR image using a series of inverted residual blocks, wherein transforming the at least one LR image to the HR image comprises:

extracting shallow features from the set of downsampled sub-images, and

processing the shallow features using the series of inverted residual blocks to generate enhanced features corresponding to the set of downsampled sub-images, wherein the shallow features are gradually modified by the series of inverted residual blocks; and

facilitating, by the video processing system, display of the HR video on a display screen of an electronic device associated with a content viewer.

20. The method as claimed in claim 19 , wherein each inverted residual block is configured to perform at least a first convolution and a second convolution to transform the feature space corresponding to the set of downsampled sub-images from a low-dimensional feature space to an enhanced feature space, wherein the enhanced feature space comprises the enhanced features corresponding to the set of downsampled sub-images.