Training classifiers for deblurring images
View Patent ↗A classifier training system trains a classifier for evaluating image deblurring quality using a set of scored deblurred images. In some embodiments, the classifier training system trains the classifier based on a number of sub-images extracted from the scored deblurred images. An image deblurring system applies a number of different deblurring transformations to a given blurry reference image and uses the classifier trained by the classifier training system to evaluate deblurring quality, thereby finding a highest-quality deblurred image. In some embodiments, the classifier training system trains the classifier in the frequency domain, and the image deblurring system uses the classifier trained by the classifier training system to evaluate deblurring quality in the frequency domain. In some embodiments, the image deblurring system applies the different deblurring transformations iteratively.
1. A machine-readable non-transitory storage medium encoded with instructions that, when executed by a processor, cause the processor to perform a method comprising:
generating a plurality of perturbed deblurring kernels by perturbing a first deblurring kernel;
applying each of the plurality of perturbed deblurring kernels to a blurry image to produce a plurality of deblurred images;
determining a plurality of deblurring scores by applying a classifier to each of the plurality of deblurred images;
selecting, using the plurality of deblurring scores, one of the plurality of perturbed deblurring kernels, as the selected deblurring kernel;
applying a frequency domain transform to the blurry image; and
applying the selected deblurring kernel to the blurry image to produce a deblurred image.
2. The medium of claim 1 , wherein plurality of perturbed deblurring kernels are generated using simulated annealing with varying parameters for each of the perturbed deblurring kernels.
3. A machine-readable non-transitory storage medium encoded with instructions that, when executed by a processor, cause the processor to perform a method comprising:
generating a first perturbed deblurring kernel by perturbing a first deblurring kernel;
generating a second perturbed deblurring kernel by perturbing one of the first deblurring kernel and the first perturbed deblurring kernel;
applying the first perturbed deblurring kernel to a blurry image to produce a first deblurred image;
applying the second perturbed deblurring kernel to the blurry image to produce a second deblurred image;
determining a first deblurring score by applying a classifier to the first deblurred image;
determining a second deblurring score by applying the classifier to the second deblurred image; and
selecting, using the first deblurring score and the second deblurring score, one of the first perturbed deblurring kernel and the second perturbed deblurring kernel as the deblurring kernel.
4. The medium of claim 3 , further comprising applying the deblurring kernel to the blurry image to produce a deblurred image.
5. The medium of claim 3 , wherein first perturbed deblurring kernel and the second perturbed deblurring kernel are generated using simulated annealing with varying parameters for each of the perturbed deblurring kernels.
6. The medium of claim 3 , further comprising:
generating a third perturbed deblurring kernel based on one or more of the first deblurring kernel, the first perturbed deblurring kernel, and the second perturbed deblurring kernel; and
determining a third deblurring score by applying the classifier to the second deblurred image.
7. The medium of claim 6 , further comprising:
selecting, using the first deblurring score, the second deblurring score, and the third deblurring score, one of the first perturbed deblurring kernel, the second perturbed deblurring kernel, and the third perturbed deblurring kernel as the deblurring kernel.
8. The medium of claim 3 , wherein identifying a first deblurring kernel based on a blurry image additionally comprises applying a frequency domain transform to the blurry image.
9. The medium of claim 3 , wherein generating the first perturbed deblurring kernel comprises applying a first transform to the first deblurring kernel, and wherein generating the second perturbed deblurring kernel comprises applying a second transform to one of the first deblurring kernel and the first perturbed deblurring kernel.
10. The medium of claim 3 , wherein applying the first perturbed deblurring kernel to the blurry image to produce a first deblurred image comprises:
applying a frequency domain transform to the blurry image;
applying a frequency domain transform to the first perturbed deblurring kernel; and
producing the first deblurred image in the frequency domain.
11. The medium of claim 3 , wherein the classifier is trained by:
applying a plurality of deblur transformations to a plurality of blurry references images thereby producing a plurality of deblurred images;
scoring each of the plurality of deblurred images;
determining an image feature vector characterizing one or more visual aspects of the image; and
training the classifier based on the image feature vectors from the deblurred images and the corresponding scores of the deblurred images.
12. A computer-implemented method, the method comprising:
generating a first perturbed deblurring kernel by perturbing a first deblurring kernel;
generating a second perturbed deblurring kernel by perturbing one of the first deblurring kernel and the first perturbed deblurring kernel;
applying the first perturbed deblurring kernel to the blurry image to produce a first deblurred image;
applying the second perturbed deblurring kernel to the blurry image to produce a second deblurred image;
determining a first deblurring score by applying a classifier to the first deblurred image;
determining a second deblurring score by applying the classifier to the second deblurred image; and
selecting, using the first deblurring score and the second deblurring score, one of the first perturbed deblurring kernel and the second perturbed deblurring kernel as the deblurring kernel.
13. The method of claim 12 , further comprising applying the deblurring kernel to the blurry image to produce a deblurred image.
14. The method of claim 12 , wherein first perturbed deblurring kernel and the second perturbed deblurring kernel are generated using simulated annealing with varying parameters for each of the perturbed deblurring kernels.
15. The method of claim 12 , further comprising:
generating a third perturbed deblurring kernel based on one or more of the first deblurring kernel, the first perturbed deblurring kernel, and the second perturbed deblurring kernel; and
determining a third deblurring score by applying the classifier to the second deblurred image.
16. The method of claim 15 , further comprising:
selecting, using the first deblurring score, the second deblurring score, and the third deblurring score, one of the first perturbed deblurring kernel, the second perturbed deblurring kernel, and the third perturbed deblurring kernel as the deblurring kernel.
17. The method of claim 12 , wherein identifying a first deblurring kernel based on a blurry image additionally comprises applying a frequency domain transform to the blurry Image.
18. The method of claim 12 , wherein generating the first perturbed deblurring kernel comprises applying a first transform to the first deblurring kernel, and wherein generating the second perturbed deblurring kernel comprises applying a second transform to the first deblurring kernel.
19. The method of claim 12 , wherein applying the first perturbed deblurring kernel to the blurry image to produce a first deblurred image comprises:
applying a frequency domain transform to the blurry image;
applying a frequency domain transform to the first perturbed deblurring kernel; and
producing the first deblurred image in the frequency domain.
20. The method of claim 12 , wherein the classifier is trained by:
applying a plurality of deblur transformations to a plurality of blurry references images thereby producing a plurality of deblurred images;
scoring each of the plurality of deblurred images;
determining an image feature vector characterizing one or more visual aspects of the image; and
training the classifier based on the image feature vectors from the deblurred images and the corresponding scores of the deblurred images.