Machine learning-based approaches for synthetic training data generation and image sharpening
A method includes obtaining an input image that contains blur. The method also includes providing the input image to a trained machine learning model, where the trained machine learning model includes (i) a shallow feature extractor configured to extract one or more feature maps from the input image and (ii) a deep feature extractor configured to extract deep features from the one or more feature maps. The method further includes using the trained machine learning model to generate a sharpened output image. The trained machine learning model is trained using ground truth training images and input training images, where the input training images include versions of the ground truth training images with blur created using demosaic and noise filtering operations.
1. A method comprising:
obtaining an input image that contains blur;
providing the input image to a trained machine learning model, the trained machine learning model comprising (i) a shallow feature extractor configured to extract one or more feature maps from the input image and (ii) a deep feature extractor configured to extract deep features from the one or more feature maps; and
using the trained machine learning model to generate a sharpened output image;
wherein the trained machine learning model is trained using ground truth training images and input training images, the input training images comprising versions of the ground truth training images with blur created using demosaic and noise filtering operations.
2. The method of claim 1 , further comprising:
training a machine learning model using the ground truth training images and the input training images in order to generate the trained machine learning model.
3. The method of claim 2 , further comprising:
generating the input training images by:
introducing noise into image data used to form the ground truth training images using a synthetic noise model in order to generate noisy images;
performing a demosaic interpolation using the noisy images in order to generate demosaiced images; and
applying a noise filter to the demosaiced images in order to generate the input training images.
4. The method of claim 1 , wherein:
the shallow feature extractor comprises a convolution layer configured to extract the one or more feature maps; and
the deep feature extractor comprises multiple residual convolutional layers and multiple channel attention layers configured to generate one or more improved feature maps, the one or more improved feature maps containing the deep features.
5. The method of claim 1 , wherein:
the deep feature extractor comprises multiple residual groups; and
each residual group comprises multiple residual channel attention blocks (RCABs).
6. The method of claim 1 , wherein the trained machine learning model represents a single-stage machine learning model.
7. The method of claim 1 , wherein the trained machine learning model comprises a raw image enhancement network.
8. An apparatus comprising:
at least one processing device configured to:
obtain an input image that contains blur;
process the input image using a trained machine learning model, the trained machine learning model comprising (i) a shallow feature extractor configured to extract one or more feature maps from the input image and (ii) a deep feature extractor configured to extract deep features from the one or more feature maps; and
use the trained machine learning model to generate a sharpened output image;
wherein the trained machine learning model is trained using ground truth training images and input training images, the input training images comprising versions of the ground truth training images with blur created using demosaic and noise filtering operations.
9. The apparatus of claim 8 , wherein the at least one processing device is further configured to train a machine learning model using the ground truth training images and the input training images in order to generate the trained machine learning model.
10. The apparatus of claim 9 , wherein:
the at least one processing device is further configured to generate the input training images; and
to generate the input training images, the at least one processing device is configured to:
introduce noise into image data used to form the ground truth training images using a synthetic noise model in order to generate noisy images;
perform a demosaic interpolation using the noisy images in order to generate demosaiced images; and
apply a noise filter to the demosaiced images in order to generate the input training images.
11. The apparatus of claim 8 , wherein:
the shallow feature extractor comprises a convolution layer configured to extract the one or more feature maps; and
the deep feature extractor comprises multiple residual convolutional layers and multiple channel attention layers configured to generate one or more improved feature maps, the one or more improved feature maps containing the deep features.
12. The apparatus of claim 8 , wherein:
the deep feature extractor comprises multiple residual groups; and
each residual group comprises multiple residual channel attention blocks (RCABs).
13. The apparatus of claim 8 , wherein the trained machine learning model represents a single-stage machine learning model.
14. The apparatus of claim 8 , wherein the trained machine learning model comprises a raw image enhancement network.
15. A non-transitory computer readable medium containing instructions that when executed cause at least one processor to:
obtain an input image that contains blur;
process the input image using a trained machine learning model, the trained machine learning model comprising (i) a shallow feature extractor configured to extract one or more feature maps from the input image and (ii) a deep feature extractor configured to extract deep features from the one or more feature maps; and
use the trained machine learning model to generate a sharpened output image;
wherein the trained machine learning model is trained using ground truth training images and input training images, the input training images comprising versions of the ground truth training images with blur created using demosaic and noise filtering operations.
16. The non-transitory computer readable medium of claim 15 , wherein the instructions when executed further cause the at least one processor to train a machine learning model using the ground truth training images and the input training images in order to generate the trained machine learning model.
17. The non-transitory computer readable medium of claim 16 , wherein:
the instructions when executed further cause the at least one processor to generate the input training images; and
the instructions that when executed cause the at least one processor to generate the input training images comprise instructions that when executed cause the at least one processor to:
introduce noise into image data used to form the ground truth training images using a synthetic noise model in order to generate noisy images;
perform a demosaic interpolation using the noisy images in order to generate demosaiced images; and
apply a noise filter to the demosaiced images in order to generate the input training images.
18. The non-transitory computer readable medium of claim 15 , wherein:
the shallow feature extractor comprises a convolution layer configured to extract the one or more feature maps; and
the deep feature extractor comprises multiple residual convolutional layers and multiple channel attention layers configured to generate one or more improved feature maps, the one or more improved feature maps containing the deep features.
19. The non-transitory computer readable medium of claim 15 , wherein:
the deep feature extractor comprises multiple residual groups; and
each residual group comprises multiple residual channel attention blocks (RCABs).
20. The non-transitory computer readable medium of claim 15 , wherein the trained machine learning model represents a single-stage machine learning model.
21. A method comprising:
obtaining multiple ground truth training images;
generating multiple input training images using the ground truth training images, wherein generating the input training images comprises performing demosaic and noise filtering operations to add blur to the ground truth training images; and
training a machine learning model to remove blur from images and generate sharpened images using the ground truth training images and the input training images.
22. The method of claim 21 , wherein:
each ground truth training image has a first ISO value; and
each input training image is generated using an associated one of the ground truth training images and has a second ISO value larger than the first ISO value of the associated ground truth training image.
23. The method of claim 21 , wherein performing the demosaic and noise filtering operations comprises:
introducing noise into image data used to form the ground truth training images using a synthetic noise model in order to generate noisy images;
performing a demosaic interpolation using the noisy images in order to generate demosaiced images; and
applying a noise filter to the demosaiced images in order to generate the input training images.
24. The method of claim 23 , wherein the synthetic noise model models read and shot noise.
25. The method of claim 23 , wherein:
the demosaic interpolation converts raw image data into YUV image data; and
the noise filter is applied to the YUV image data in order to generate raw RGB image data.
26. The method of claim 21 , wherein training the machine learning model comprises training the machine learning model based on an L1 Charbonnier with edge loss using an Adam optimizer.