Joint rolling shutter correction and image deblurring
A method for jointly removing rolling shutter (RS) distortions and blur artifacts in a single input RS and blurred image is presented. The method includes generating a plurality of RS blurred images from a camera, synthesizing RS blurred images from a set of GS sharp images, corresponding GS sharp depth maps, and synthesized RS camera motions by employing a structure-and-motion-aware RS distortion and blur rendering module to generate training data to train a single-view joint RS correction and deblurring convolutional neural network (CNN), and predicting an RS rectified and deblurred image from the single input RS and blurred image by employing the single-view joint RS correction and deblurring CNN.
1. A computer-implemented method executed on a processor for jointly removing rolling shutter (RS) distortions and blur artifacts in a single input RS and blurred image, the method comprising:
generating a plurality of RS blurred images from a camera;
synthesizing RS blurred images from a set of GS sharp images, corresponding GS sharp depth maps, and synthesized RS camera motions by employing a structure-and-motion-aware RS distortion and blur rendering module to generate training data to train a single-view joint RS correction and deblurring convolutional neural network (CNN); and
predicting an RS rectified and deblurred image from the single input RS and blurred image by employing the single-view joint RS correction and deblurring CNN;
wherein single-view joint RS correction and deblurring is performed in a single forward pass.
2. The method of claim 1 , further comprising, during a training phase, employing the structure-and-motion-aware RS distortion and blur rendering module with the set of GS sharp images, the corresponding GS sharp depth maps, and the synthesized RS camera motions to generate synthesized RS blurred images to train the single-view joint RS correction and deblurring convolutional neural network (CNN).
3. The method of claim 1 , further comprising, during a testing phase, providing the single RS and blurred image to the single-view joint RS correction and deblurring CNN to produce corresponding RS rectified and deblurred images in a single forward pass.
4. The method of claim 2 , wherein a generator network predicts corresponding GS sharp images.
5. The method of claim 2 , wherein a discriminator network takes either a predicted GS sharp image or a ground truth GS sharp image to classify such image as either “real” or “fake”.
6. The method of claim 5 , wherein the generator network and the discriminator network are jointly trained via an adversarial loss.
7. The method of claim 6 , wherein a perceptual loss is added to minimize differences between activations of different layers of a visual geometry group (VGG) network.
8. The method of claim 6 , wherein the generator network includes an encoder and a decoder, the encoder extracting shared features between single-view RS correction and deblurring from the single input RS and blurred image into a compact bottleneck, and the decoder upsampling the compact bottleneck to an original input image size.
9. A non-transitory computer-readable storage medium comprising a computer-readable program for jointly removing rolling shutter (RS) distortions and blur artifacts in a single input RS and blurred image, wherein the computer-readable program when executed on a computer causes the computer to perform the steps of:
generating a plurality of RS blurred images from a camera;
synthesizing RS blurred images from a set of GS sharp images, corresponding GS sharp depth maps, and synthesized RS camera motions by employing a structure-and-motion-aware RS distortion and blur rendering module to generate training data to train a single-view joint RS correction and deblurring convolutional neural network (CNN); and
predicting an RS rectified and deblurred image from the single input RS and blurred image by employing the single-view joint RS correction and deblurring CNN;
wherein single-view joint RS correction and deblurring is performed in a single forward pass.
10. The non-transitory computer-readable storage medium of claim 9 , wherein, during a training phase, a set of GS sharp images, corresponding GS sharp depth maps, and synthesized RS camera motions are employed by the structure-and-motion-aware RS distortion and blur rendering module to generate synthesized RS blurred images to train the single-view joint RS correction and deblurring convolutional neural network (CNN).
11. The non-transitory computer-readable storage medium of claim 9 , wherein, during a testing phase, the single RS and blurred image is provided to the single-view joint RS correction and deblurring CNN to produce corresponding RS rectified and deblurred images in a single forward pass.
12. The non-transitory computer-readable storage medium of claim 10 , wherein a generator network predicts corresponding GS sharp images.
13. The non-transitory computer-readable storage medium of claim 10 , wherein a discriminator network takes either a predicted GS sharp image or a ground truth GS sharp image to classify such image as either “real” or “fake”.
14. The non-transitory computer-readable storage medium of claim 13 , wherein the generator network and the discriminator network are jointly trained via an adversarial loss.
15. The non-transitory computer-readable storage medium of claim 14 , wherein a perceptual loss is added to minimize differences between activations of different layers of a visual geometry group (VGG) network.
16. The non-transitory computer-readable storage medium of claim 14 , wherein the generator network includes an encoder and a decoder, the encoder extracting shared features between single-view RS correction and deblurring from the single input RS and blurred image into a compact bottleneck, and the decoder upsampling the compact bottleneck to an original input image size.
17. A system for jointly removing rolling shutter (RS) distortions and blur artifacts in a single input RS and blurred image, the system comprising:
a memory; and
one or more processors in communication with the memory configured to:
generate a plurality of RS blurred images from a camera;
synthesize RS blurred images from a set of OS sharp images, corresponding GS sharp depth maps, and synthesized RS camera motions by employing a structure-and-motion-aware RS distortion and blur rendering module to generate training data to train a single-view joint RS correction and deblurring convolutional neural network (CNN); and
predict an RS rectified and deblurred image from the single input RS and blurred image by employing the single-view joint RS correction and deblurring CNN;
wherein single-view joint RS correction and deblurring is performed in a single forward pass.
18. The system of claim 17 ,
wherein a generator network predicts corresponding GS sharp images and a discriminator network takes either a predicted GS sharp image or a ground truth GS sharp image to classify such image as either “real” or “fake,”
wherein the generator network and the discriminator network are jointly trained via an adversarial loss, and
wherein a perceptual loss is added to minimize differences between activations of different layers of a visual geometry group (VGG) network.