Image hole filling that accounts for global structure and local texture

Patent metadata

Field	Value
Publication number	US-10290085-B2
Application number	US-201615379337-A
Country	US
Kind code	B2
Filing date	Dec 14, 2016
Priority date	Dec 14, 2016
Publication date	May 14, 2019
Grant date	May 14, 2019

Abstract

Image hole filling that account for global structure and local texture. One exemplary technique involves using both a content neural network and a texture neural network. The content neural network is trained to encode image features based on non-hole image portions and decode the image features to fill holes. The texture neural network is trained to extract image patch features that represent texture. The exemplary technique receives an input image that has a hole and uses the two neural networks to fill the hole and provide a result image. This is accomplished by selecting pixel values for the hole based on a content constraint that uses the content neural network to account for global structure and a texture constraint that uses the texture neural network to account for local texture. For example, the pixel values can be selected by optimizing a loss function that implements the constraints.

First claim

What is claimed is:

1. A method, performed by a computing device, for hole filling in images that accounts for global structure and local texture, the method comprising:
receiving an input image, the input image having a hole, the hole comprising a portion of the input image having unknown, removed, or missing content;
selecting content to fill the hole, wherein the content to fill the hole is selected based on a content constraint and a texture constraint,
wherein selecting the content based on the content constraint comprises selecting content that is consistent with a global structure of the image using a content neural network, the content neural network identifying the global structure of the image, and
wherein selecting the content based on the texture constraint comprises selecting content with texture similar to the texture of the image outside the hole using a texture neural network, the texture neural network identifying image features that are used to assess texture similarity; and
providing a result image by filling in the hole of the input image based on the content selected to fill the hole.

2. The method of claim 1 , wherein selecting the content to fill the hole comprises optimizing a loss function, the loss function comprising a content term that applies the content constraint and a texture term that applies the texture constraint.

3. The method of claim 2 ,
wherein the content neural network encodes high level features of the image representing the global structure of the image and decodes the high level features to produce an initial predication of content to fill the hole that is consistent with the global structure of the image; and
wherein the content term penalizing selecting content to fill the hole that deviates from the initial prediction of content to fill the hole.

4. The method of claim 2 ,
wherein the texture neural network is a classification network that identifies different image features using feature maps within a layer of the classification network that encodes mid-level image information, and
wherein the texture term penalizing selecting content to fill the hole that includes patches with feature maps that differ from feature maps of patches outside of the hole.

5. The method of claim 2 , wherein optimizing the loss function comprises using a backpropagation technique, wherein the backpropagation technique iteratively solves the loss function using an output prediction of the content to fill the hole from each iteration as an input for a next iteration.

6. The method of claim 5 , wherein optimizing the loss function comprises determining the input for a first iteration using the content neural network on the input image with a mean value of a collection of natural images to initialize the hole.

7. The method of claim 2 , wherein optimizing the loss function comprises iteratively solving the loss function,
wherein the content constraint is applied to penalize deviation between a forward prediction of the content neural network and a previous prediction of the content neural network, and
wherein the texture constraint is applied to penalize deviations of patches within the hole to patches outside the hole that are identified as being most similar, wherein the deviations are determined by comparing extracted image features that represent texture using a layer of the texture neural network that encodes mid-level image information.

8. The method of claim 1 further comprising enhancing the content selected to fill the hole using a patch match technique, wherein the content selected for the hole is input to the patch match technique, wherein the patch match technique adjusts pixels of patches inside the hole to match pixels of patches outside the hole.

9. The method of claim 1 further comprising training the content neural network to predict pixels for holes that approximate original content in the training images.

10. The method of claim 1 further comprising training the texture neural network for image classification, wherein a layer of the texture neural network encodes mid-level image information used to represent texture.

11. A computing device comprising:
a means for receiving an input image having a hole, the hole comprising a portion of the input image having unknown, removed, or missing content; and
a means for filling the hole of the input image using a content neural network to account for global content and a texture neural network to account for local texture, wherein the means for filling the hole comprises a processor executing instructions to optimize a loss function, the loss function comprising a content term that uses a content neural network and a texture term that uses the texture neural network.

12. The computing device of claim 11 further comprising a means for training the content neural network, the content neural network trained to encode image features based on non-hole image portions and decode the image features to fill holes.

13. The computing device of claim 11 , further comprising a means for training the texture neural network, the texture neural network trained to extract image patch features that represent texture.

14. The computing device of claim 11 , wherein the means for filling the hole comprises a means for selecting content to fill the hole based on a content constraint that uses the content neural network and a texture constraint that uses the texture neural network.

15. The computing device of claim 11 further comprising a means for performing a patch match technique to add high frequency texture within the hole.

16. A non-transitory computer-readable medium storing instructions, the instructions comprising instructions for:
receiving an input image, the input image having a hole, the hole comprising a portion of the input image having unknown, removed, or missing content;
selecting content to fill the hole based on a content constraint that uses a content neural network to account for global structure and a local texture constraint that uses a texture neural network to account for local texture including optimizing a loss function, the loss function comprising a content term applying the content constraint and a texture term applying the texture constraint; and
filling in the hole of the input image based on the content selected to fill the hole.

17. The non-transitory computer-readable medium of claim 16 , wherein optimizing the loss function comprises using a backpropagation technique, wherein the backpropagation technique iteratively solves the loss function using an output from each iteration as an input for a next iteration.

18. The non-transitory computer-readable medium of claim 17 , wherein iteratively solving the loss function comprises:
applying the content constraint by penalizing deviation between a forward prediction of the content neural network and a previous prediction of the content neural network; and
applying the texture constraint by penalizing deviations of patches within the hole to patches outside the hole that are identified as being most similar, wherein the deviations are determined by comparing extracted image features that represent texture using a layer of the texture neural network that encodes mid-level image information.

Assignees

• Adobe Inc

Inventors

• Lin Zhe
• Lu Xin
• Yang Chao

Classifications

• G06V10/82

  using neural networks · CPC title

• G06V10/764

  using classification, e.g. of video objects · CPC title

• G06N3/045

  Combinations of networks · CPC title

• G06F18/24

  Classification techniques · CPC title

• G06N3/0455

  Auto-encoder networks; Encoder-decoder networks · CPC title