System and method for burst image restoration and enhancement

1. A mobile device, comprising:

at least one camera having an image capture circuit operating in a mode to capture a RAW image burst containing ten or more RGB images; and

processing circuitry, including a neural network engine, to generate a single enhanced image from the RAW image burst, the neural network engine executing program instructions including

an edge boosting feature alignment stage to remove inter-frame spatial and color misalignment from the RAW image burst to obtain a plurality of aligned burst frames;

a pseudo-burst feature fusion stage to create a set of pseudo-burst features that combine complementary information from the aligned burst frames; and

an adaptive group upsampling stage to progressively increase spatial resolution while merging the set of pseudo-burst features and output the single enhanced image.

2. The mobile device of claim 1 , wherein the edge boosting feature alignment stage performs feature processing to reduce noise in initial burst features followed by alignment of all images in the RAW image burst with a base frame, wherein the base frame is one of the images in the RAW image burst.

3. The mobile device of claim 2 , wherein the feature processing includes at least one residual global context attention block for obtaining global features of the RAW image burst among feature channels.

4. The mobile device of claim 1 , wherein the edge boosting feature alignment stage further performs alignment correction by computing a high-frequency residue after determining a difference between refined aligned features and base frame features.

5. The mobile device of claim 1 , wherein the pseudo-burst feature fusion stage includes generation of feature tensors obtained by concatenating corresponding channel-wise features extracted from all aligned burst frames.

6. The mobile device of claim 2 , wherein the pseudo-burst feature fusion stage includes a light-weight U-Net to extract multi-scale features, in which the light-weight U-net uses the feature processing of the edge boosting feature alignment stage.

7. The mobile device of claim 1 , wherein the adaptive group upsampling stage takes as input the set of pseudo-burst features produced by the pseudo-burst feature fusion stage and outputs a super-resolved output via three-level progressive upsampling.

8. The mobile device of claim 1 , wherein the camera is a smartphone camera, and the neural network engine generates a single super-resolution image from a lower resolution RAW image burst.

9. The mobile device of claim 1 , wherein the camera is a smartphone camera, and the neural network engine generates a well-light RGB image from low-light RAW image burst.

10. The mobile device of claim 1 , wherein the at least one camera is a smartphone camera,

wherein the adaptive group upsampling stage includes a group convolution layer, and

wherein the neural network engine generates a noise-free image from a noisy RGB image burst.

11. A non-transitory computer readable storage medium storing a mobile application having program instructions, which when executed in a neural network engine, performs a method comprising:

removing inter-frame spatial and color misalignment from a RAW image burst to obtain a plurality of aligned burst frames;

creating a set of pseudo-burst features that combine complementary information from the aligned burst frames; and

progressively increasing spatial resolution while merging the set of pseudo-burst features and outputting a single enhanced image.

12. The non-transitory computer readable storage medium of claim 11 , further comprising:

performing feature processing to reduce noise in initial burst features followed by performing alignment of all images in the RAW image burst with a base frame, wherein the base frame is one of the images in the RAW image burst.

13. The non-transitory computer readable storage medium of claim 12 , further comprising:

obtaining global features of the RAW image burst among feature channels.

14. The non-transitory computer readable storage medium of claim 11 , wherein the removing to obtain aligned burst frames further comprises:

performing alignment correction by computing a high-frequency residue after determining a difference between refined aligned features and base frame features.

15. The non-transitory computer readable storage medium of claim 11 , further comprising:

generating feature tensors obtained by concatenating corresponding channel-wise features extracted from all aligned burst frames.

16. The non-transitory computer readable storage medium of claim 12 , further comprising:

a light-weight U-Net to extract multi-scale features, in which the light-weight U-net uses the feature processing.

17. The non-transitory computer readable storage medium of claim 11 , further comprising:

receiving the set of pseudo-burst features and outputting a super-resolved output via three-level progressive upsampling.

18. The non-transitory computer readable storage medium of claim 11 , wherein the neural network engine generates a single super-resolution image from a lower resolution RAW image burst.

19. The non-transitory computer readable storage medium of claim 11 , wherein the neural network engine generates a well-light RGB image from low-light RAW image burst.

20. The non-transitory computer readable storage medium of claim 11 , wherein the neural network engine generates a noise-free image from a noisy RGB image burst.