System, method and apparatus for scene recognition

1. An image processing system comprising:

an image processing computer coupled to a wide area network and to an image source computer; and

one or more processors configured to perform actions, including to:

generate an image distance metric from a set of images, wherein the image distance metric indicates image feature weights corresponding to image features associated with the set of images;

retrieve an input image;

extract a first set of input image features from the input image, wherein

the first set of input image features corresponds to the image distance metric;

extract a set of image features from individual ones of images in the set of images, wherein set of image features corresponds to the image distance metric;

using the image distance metric, compute one or more image feature distances between the first set of input image features and individual ones of the set of image features;

select, from the set of images, a subset of images based at least in part on image feature distances, the subset of images corresponding to a set of scene types;

determine a scene type from the set of scene types; and

assign the scene type to the input image.

2. The image processing system of claim 1 , wherein the image distance metric indicates a second set of image features and a set of image feature weights corresponding to the second set of image features.

3. The image processing system of claim 1 , wherein the actions further include to retrieve a source scene image via a network interface and segment the source scene image into multiple images, wherein the multiple images include the input image.

4. The image processing system of claim 3 , wherein the actions further include to:

recognize scene types of one or more of the multiple images; and

based at least in part on the scene types, determine a scene type for the source scene image.

5. The image processing system of claim 1 , wherein: the set of input image features includes an estimated depth of the input image; and the first set of image features for individual ones of the images in the set of images includes an estimated depth of the image.

6. The image processing system of claim 1 , wherein the actions further include to send the image distance metric to a client computer.

7. The image processing system of claim 6 , wherein the actions further include to send the first set of image features for individual ones of the images in the set of images to the client computer.

8. The image processing system of claim 1 , wherein the actions further include to receive the first set of image features for the input image from a client computer.

9. The image processing system of claim 1 , wherein the actions further include to retrieve the input image via a network interface coupled to the image processing computer.

10. The image processing system of claim 1 , wherein the image processing system is a distributed computing system.

11. The image processing system of claim 10 , wherein the image processing system is a cloud computing system.

12. The image processing system of claim 1 , wherein the actions further include to:

extract a set of raw image features for individual ones of the images in the set of images; and

reduce dimensionality of the sets of raw image features to derive the image features and the image feature weights.

13. The image processing system of claim 12 , wherein the actions further include to:

perform scene understanding on an unlabeled image to match a scene type to the unlabeled image;

determine that the scene type is accurate;

label the unlabeled image with the scene type to generate a labeled image; and

generate a refined image distance metric from the set of images and the labeled image.

14. The image processing system of claim 1 , wherein the actions further include to:

based at least in part on the image distance metric, learn a classification model from a plurality of images of each scene type in a list of scene types;

apply the classification model to the input image to generate a set of matching scores; and

based on the set of matching scores, select the subset of images.

15. A method comprising:

generating an image distance metric from a set of images, wherein the image distance metric indicates image feature weights corresponding to image features;

retrieving an input image;

extracting a first set of input image features from the input image, wherein the first set of input image features corresponds to the image distance metric;

extracting a set of image features from each image in the set of images, wherein extracted sets of image features corresponds to the image distance metric;

computing, based at least in part on the image distance metric, one or more image feature distances between the first set of input image features and individual ones of the extracted sets of image features;

selecting, from the set of images, a subset of images based at least in part on the one or more image feature distances, the subset of images corresponding to a set of scene types;

determining a scene type from the set of scene types; and

assigning the scene type to the input image.

16. The method of claim 15 , wherein the image distance metric indicates a second set of image features and a set of image feature weights corresponding to the second set of image features.

17. The method of claim 16 , further comprising retrieving a source scene image via a network interface and segmenting the source scene image into multiple images.

18. The method of claim 17 , further comprising:

recognizing scene types of one or more of the multiple images; and

determining, based at least in part on the scene types, a scene type for the source scene image.

19. The method of claim 16 , wherein: the set of input image features includes an estimated depth of the input image and the first set of image features for individual ones of the images in the set of images includes an estimated depth of the image.

20. The method of claim 17 , further comprising,

learning, based at least in part on the image distance metric, a classification model from a plurality of images of individual scene types in a list of scene types;

applying the classification model to the input image to generate a set of matching scores; and

selecting, based at least in part on the set of matching scores, the subset of images.