Image summarization by a learning approach
View Patent ↗An automated system and a method for extracting a region of interest from a digital image are disclosed. The method includes identifying a subset of training images from a larger set of training images, each training image in the set having a respective identified region of interest. The subset is identified based on a measure of similarity between the digital image and the images in the set of training images. At least one region of interest is extracted from the digital image based on an analysis of the identified regions of interest in the subset of training images.
1. A method for extracting a region of interest from a digital image, the method comprising:
from a set of training images, each training image in the set having a respective identified region of interest, with a processor, identifying a subset of similar training images, based on a measure of similarity between the digital image and the images in the set of training images, the identifying of the subset of similar training images comprising identifying the K nearest neighboring images, wherein K has a value of at least 10; and
extracting a region of interest for the digital image based on the identified regions of interest in the subset of training images.
2. The method of claim 1 , wherein the extracting of the region of interest for the digital image comprises:
for each of a set of pixels of the digital image, computing a probability of being in a region of interest, based on the identified regions of interest of the similar training images; and
extracting a region of interest based on the computed probability.
3. The method of claim 1 , wherein the extracting of a region of interest for the digital image is also based on a saliency determination.
4. The method of claim 1 , wherein the identified region of interest of each training image in the set of training images is described by a parameter vector of the general form =[x c (i),y c (i),w(i),h(i),z(i)], where (x c (i),y c (i)) represent the region of interest's center of mass, h(i) represents the height of the region of interest, w(i) represents the width of the region of interest, and z(i) represents a rescaling factor.
5. The method of claim 1 , wherein the identifying of the subset of similar training images comprises extracting features from the digital image, deriving a representation of the image based on the extracted features, and comparing the representation with representations generated for the training images.
6. The method of claim 1 , wherein the identifying of the subset of similar training images is based on at least one of the group consisting of spectral information, color, structural similarity, and content similarity.
7. The method of claim 1 , wherein the number of training images in the set of training images is at least 20K.
8. The method of claim 1 , wherein the computing the probability of a region of interest for the digital image comprises computing a probability map in which a probability is assigned to each pixel of the image of being in a region of interest.
9. The method of claim 8 , wherein the probability of each pixel of the original image of being in a region of interest is a function of an optionally weighted sum of a set of Gaussian functions centered on the regions of interest of the similar images.
10. The method of claim 1 , wherein the extraction includes extracting a region which contains all the pixels having at least a threshold probability of being in a region of interest.
11. The method of claim 10 , wherein the extracted region has a preselected ratio of height to width.
12. The method of claim 10 , wherein the threshold is selected such that the region has an area which is at least equal to a mean area of the regions of interest of the similar images.
13. The method of claim 1 , further comprising registering the similar images and modifying reference coordinates of the respective identified regions of interest based on the registration.
14. The method of claim 1 , further comprising:
prior to the identifying of the subset of similar training images:
for each image of a set of training images, manually identifying a region of interest within the image and storing reference coordinates of the region of interest.
15. The method of claim 1 , wherein the extracting the region of interest for the digital image comprises extracting a first region of interest from the digital image, the method further comprising:
extracting a second region of interest for at least a second digital image based on identified regions of interest in a second subset of the training images; and
storing the first and second regions of interest.
16. A computer program product embodied on a non-transitory computer-readable medium encoding instructions which, when executed by a computer, perform the method of claim 1 .
17. A variable printing method comprising:
forming a first document by combining image data with a first region of interest extracted from a first image by the method of claim 1 ;
separately, forming a second document by combining the same image data with a second region of interest extracted from a second image by the method of claim 1 ; and
printing the first and second documents.
18. A system for extracting a region of interest from a digital image, the system comprising:
memory which stores instructions for performing the method of claim 1 ; and
a processor, in communication with the memory, for executing the instructions.
19. A computer implemented system for extracting a region of interest from a digital image, the system comprising:
memory which stores a set of training images, each training image in the set being associated with a respective identified region of interest;
an image identification component for identifying a subset of the training images, based on a computed similarity measure between the digital image and the images in the set of training images, the identifying of the subset of similar training images comprising identifying the K nearest neighboring images, wherein K has a value of at least 10;
a probability estimator which, for pixels of the digital image, estimates the probability of being in a region of interest, based on the identified regions of interest of the subset of training images;
a region of interest extractor which extracts a region of interest for the digital image based on estimated probabilities; and
a processor which implements the image identification component, probability estimator and region of interest extractor.
20. The system of claim 19 , further comprising a display for displaying the extracted region of interest.
21. The system of claim 19 , wherein the probability estimator estimates the probability of being in a region of interest for the pixels based on reference coordinates of the identified regions of interest of the similar training images.
22. A computer-implemented method for extracting a region of interest from an image, the method comprising:
providing a set of training images, each training image in the set being associated with reference coordinates of a respective manually-identified region of interest in the image;
receiving an original image for which a region of interest is to be extracted;
with a processor, for the original image, identifying a subset of K nearest neighbor training images from the set of training images; and
generating a first probability map for the original image which assigns to pixels of the image, a probability of being in a region of interest based on the reference coordinates of the regions of interest in the K nearest neighbor training images;
optionally, generating a second probability map for the original image which assigns to pixels of the image a probability of being in a region of interest based on saliency features;
extracting a region of interest for the original image based on the first probability map and optionally the second probability map.
23. A computer program product embodied on a non-transitory computer-readable medium encoding instructions which, when executed by a computer, perform the method of claim 22 .