초록 ▼

A method and system for detecting faces at different views within images that allocates the computational effort based on a prior probability associated with a view is provided. A face detection system determines whether an image contains a face using detectors that are adapted to detect faces at va

A method and system for detecting faces at different views within images that allocates the computational effort based on a prior probability associated with a view is provided. A face detection system determines whether an image contains a face using detectors that are adapted to detect faces at various views and a filter that filters out windows of the image that are provided to a detector based on a prior probability associated with the view of the detector. Each view has an associated prior probability that a face from a collection of real-life home photographs will be at that view. The face detection system allocates increasing computational effort to a detector as the prior probability of its view increases.

대표청구항 ▼

We claim: 1. A method in a computer system for detecting whether an image contains a face, the method comprising: providing a collection of images, the images having faces at various views: for each of a plurality of views, providing a prior probability that faces in images of the collection will b

We claim: 1. A method in a computer system for detecting whether an image contains a face, the method comprising: providing a collection of images, the images having faces at various views: for each of a plurality of views, providing a prior probability that faces in images of the collection will be at that view based on a percentage of the faces in the images of the collection that are at that view; providing detectors for determining whether a window of an image contains a face, each detector adapted to detect faces with a certain view; providing a filter for each detector, the filter for filtering out windows based on a the prior probability associated with the view of the detector; after the prior probabilities have been provided, receiving an image for detecting whether the image contains a face; and for each window of the received image, generating a score relating to a likelihood that the received image contains a face; and for each detector, applying the filter of the detector to filter out the window when the generated score does not satisfy a threshold score, such that the threshold score is higher for filters of detectors of views with lower prior probabilities and lower for filters of detectors of views with higher prior probabilities; and applying the detector to the window when the window has not been filtered to identify a candidate window containing a face. 2. The method of claim 1 wherein a provided filter is applied to limit the windows to which a detector is applied. 3. The method of claim 2 wherein a provided filter is a skin color filter. 4. The method of claim 1 wherein a provided filter is applied to reduce the candidate windows after a detector is applied. 5. The method of claim 1 including merging candidate windows that overlap. 6. The method of claim 1 wherein the detectors are applied to windows of various sizes. 7. The method of claim 6 wherein the prior probability associated with a detector is used to determine to which sizes of window the detector is applied. 8. The method of claim 1 wherein a detector is a chain adaptive boosting detector. 9. A computer-readable storage medium containing instructions for controlling a computer system to detect whether a first image contains an object by a method comprising: for each of a plurality of views, providing a prior probability indicating a probability that objects within images of a collection of images will be in that view; and after providing a prior probability for each view, detecting whether the first image contains an object by, dividing the first image into windows of different sizes; and for each of a plurality of views, designating windows as candidate windows based on a filter criterion that factors in the prior probability of the designating being based on a threshold that varies with the prior probability of the view such that a view with a higher prior probability has more windows designated than a view with a lower prior probability; and applying a detector for the view to the candidate windows to detect which candidate windows contain the object; and merging overlapping candidate windows that are detected to contain the object such that the detector for a view with a higher prior probability is applied to more candidate windows of the first image than a detector with a view with a lower prior probability resulting in more computational effort being allocated to detectors for views with higher prior probabilities. 10. The computer-readable storage medium of claim 9 wherein the merging factors in a prior probability of a detector that detected that the candidate window contains the object. 11. The computer-readable storage medium of claim 9 wherein the object is a face. 12. The computer-readable storage medium of claim 11 wherein the filter criterion is based on skin color. 13. The computer-readable storage medium of claim 11 wherein the prior probability associated with a detector is used to determine to which sizes of windows the detector is applied. 14. The computer-readable storage medium of claim 11 wherein a detector is a chain adaptive boosting detector. 15. A computer system for detecting whether an image contains a face, comprising: a detector for each of a plurality of views of a face; a data store having a prior probability for each of the plurality of views, a prior probability indicating a probability that faces from a collection of images will be at the view; a component that determines sizes of windows to be processed by detectors based on prior probabilities of views a component that, after sizes of the windows are determined, receives an indication of an image for detecting whether the image contains a face; a component that determines which windows of the received image of which determined sizes should have which detector applied based on the view of the detector and the prior probability of that view, such that the determining is based on a threshold that varies with the prior probability of the view; and a component for merging windows detected by the detectors to have faces wherein more computational effort is allocated to detectors for views with higher prior probabilities. 16. The computer system of claim 15 wherein the detectors are chain adaptive boosting detectors.