The present invention claims priority to, and the benefit of, Korean Patent Application No. 10-2008-0131783, filed on Dec. 22, 2008, which is incorporated herein by reference.
The present invention relates to a system for detecting a facial image, and more particularly to facial image detecting apparatus and method.
A facial image detecting technology is used in various fields since an adaptive boosting technology has been appeared in the world. The adaptive boosting technology is used in various facial image detecting and searching systems due to detection of a facial image in real time by learning a face in off-line and a high quality facial image detection.
For instance, the adaptive boosting technology is mainly used in products having a facial image detecting function such as CCTV/Digital Video Recorder (DVR), a face recognizing smart camera, a face recognizing mobile phone, a photo-booth, a face recognizing printer and a digital still camera.
Since a method of detecting a facial image requires a number of facial images, a face detection rate does not come in question. However, a post-processing such as face authorization cannot be carried out when facial image detection in a database search or from a single image such as a photograph is failed. And thus, the facial image detection requires high precision. Moreover, a method of detecting a facial image as a result of learning a face in off-line shows a different face detection rate according to whether which learning database is used.
In addition, a method of searching a candidate of a facial image using color information of an image or segmentation of an image is not suitable for detecting a facial image requiring a high face detection rate because the method is dependent on the color information and the color information is influenced by illumination.
Therefore, in a field requiring a high face detection rate like in a single image, there is no time restriction differently from a field of detecting a facial image in real time but a more precise search is required.
In view of the above, the present invention provides facial image detecting apparatus and method capable of detecting a correct face area and a coordinate of eyes.
In accordance with a first aspect of the present invention, there is provided a method of detecting a facial image, including: pre-processing an input image;
detecting a face region from the pre-processed input image and creating facial records of the detected face region; and detecting the facial image by creating coordinates of the face and eyes in the input image by using the facial records.
In accordance with a second aspect of the present invention, there is provided an apparatus for detecting a face, including: an image input unit for receiving an input image; a pre-processor for performing pre-processing to the input image; a facial record creating unit for creating the n×n pixel-sized windows with respect to the pre-processed image, detecting face regions from the windows and creating facial records with respect to the detected face regions; and
a face detector for detecting a face by creating center coordinate of face and coordinates of eyes in the input image by using the facial records.
In the facial image detecting apparatus and method robust to face pose changes for detecting a plurality of facial images, strong face and eye classifiers that are learned with adaptive boosting in off-line are used to perform the facial image detection, the eye detection, the normalization, and the face verification while an input image is reduced by a preset magnification, facial records such as center coordinate, width, and length of faces, and center coordinates of eyes which are created during the face verification are used to perform the facial image detection, so that the facial image can be exactly detected even though various face pose changes. Next, the center coordinates of face and eyes obtained from the facial record are classified into several face candidates and eye candidates by applying clustering, and respective clusters are analyzed to detect final positions of face and eyes.
The objects and features of the present invention will become apparent from the following description of embodiment in conjunction with the accompanying drawings, in which:
Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings which form a part hereof.
The image input unit 100 receives an image from which a facial image is to be detected and applies the received image to the pre-processor 102.
The pre-processor 102 converts the image received from the image input unit 100 into a black-and-white image and performs compensation of illumination to the converted black-and-white image through histogram equalization or self quotient image (SQI) process such that illumination effect can be minimized in the detection of facial image.
The facial record creating unit 104 creates n×n sized windows with respect to the compensated black-and-white image, detects a face region by inspecting whether a face is in the respective windows, and detects coordinates of eyes. In addition, the facial record creating unit 104 normalizes facial image into a preset size based on positions of the eyes to create facial records such as a center coordinate of face, coordinates of eyes, a face size and on the like.
The facial image detector 106 classifies the facial records into a face candidate group and an eye candidate group through a clustering technique and calculates the center coordinate of face and the coordinates of eyes by calculating an average coordinate of the face region and the positions of eyes from the face candidate group and the eye candidate group.
Specifically, firstly the facial image detector 106 calculates the average coordinate of the center coordinates of the face candidate group, eliminates facial records far from the average coordinate over a preset distance, and calculates an average coordinate of the center coordinates of remaining face candidates after the elimination to obtain a final center coordinate of face. Moreover, an average coordinate of eye coordinates of the eye candidate group is calculated by using the remaining facial records after eliminating the center coordinates of faces, and eye candidates far from the average coordinate over a preset distance are eliminated, an average coordinate of eye coordinates of the remaining eye candidates is calculated again to calculate a final coordinate of eyes. A size of face is calculated by extracting face size information from the respective facial records of the face candidates used in calculating the final center coordinate and averaging the face size information.
First, when an image from which detection of a facial image is required is input to the image input unit 100 in step S200, the input image is converted into a black-and-white image and compensation of illumination is performed by the pre-processor 102 in step S202. In this case, the compensation of illumination minimizes the illumination effects by using histogram equalization or SQI process.
Thereafter, facial record information is created by the facial record creating unit 104 from the image pre-processed by the pre-processor 102. In other words, the facial record creating unit 104 creates n×n pixel-sized windows from the pre-processed image, performs the face detection, eye detection, normalization and face verification from images in the windows, and creates facial records while reducing magnification of the images in step S204.
As such, the created facial records are used to detect faces by the facial image detector 106. That is, when a plurality of facial records are created, the facial image detector 106 classifies the facial records into the face candidate group and the eye candidate group by using clustering techniques such as Self Organizing Map (SOM) or K-means clustering in step S206.
At this time, when the face candidate group and the eye candidate group are classified as described above, an average coordinate of the respective face candidates is calculated, facial records containing face candidates far from the average coordinate over a preset distance are eliminated, and an average coordinate is calculated again to calculate a final center coordinate of faces. Meanwhile, a final coordinate of eyes is calculated by which an average coordinate of the respective eye candidates is obtained, eye candidates far from the average coordinate over a preset distance are eliminated, and an average coordinate is obtained again to calculate the re-obtained average coordinate as the final coordinate of eyes in step S208.
Next, the face is detected by using the final center coordinate of face and the final coordinates of eyes in step S210.
Then, the facial record creating unit 104 extracts windows of restricted portions (n×n sized pixel) of entire regions from the reduced images (step S303) and inspects whether the images of extracted windows are faces in step S304. The inspection of faces is made by a face strong classifier that is learned by an adaptive boosting algorithm and determining whether the images of extracted windows are faces. When the image of extracted window is not a face, the facial image detection is performed to the next window. When facial image is detected, the eye detection is performed from the detected facial image in step S305.
Then, the facial record creating unit 104 calculates the positions of eyes as the positions of eyes in the input image of which size is not reduced when the positions of eyes are detected, and normalizes the input image by a preset size based on the positions of eyes in the input image in step S306. That is, the image is rotated such that the positions of eyes are on the horizontal line and the image is reduced to a preset size.
Next, the facial record creating unit 104 compensates illumination with respect to the reduced image. In this case, when the image from which a face is detected is normalized, face verification is performed to the normalized image in step S307. The reason why the face verification is performed with respect to the image from which face is detected is because, although a face is detected and eyes are detected in the previous step, another portion other than the facial image may pass the face verification.
Therefore, the face verification is performed to the normalized image. The face verification may be performed by using the strong classifier learned with the adaptive boosting or relative brightness difference of face elements. That is, eyes have lower brightness than a cheek and a mouth has brightness relatively lower than a cheek. The face verification is performed by using these features.
When all procedures as described above are passed, the facial record creating unit 104 stores a coordinate of a center point of a face, width and length of the face, and the coordinates of eyes as a single facial record in step S308. These procedures are continuously performed to a final window of the image in step S309. When the procedures are performed to the final window, a next image is reduced by a preset magnification and the above procedures are repeated in step S310. When passing the above procedures, a plurality of facial records is finally created.
As described with reference to
The facial records far from the average center coordinate over a preset distance are eliminated by using the obtained average center coordinate of faces. An average center coordinate of faces of the remaining face candidates is obtained again and is calculated as the center coordinates of faces. Face size information such as width and length of a face is calculated by averaging the remaining facial records. Then, the center coordinates of eyes are calculated with the remaining facial records. Like the above-described method, the center coordinate of eyes is calculated by which an average center coordinate of respective eyes is obtained from the eye candidates, by which eye candidates far from the average center coordinate over a preset distance are eliminated, and by which an average center coordinate of the remaining eye candidates is set to a final center coordinate of the remaining eye candidates.
The above-described procedures are advantageous in that the correct centers of faces and the correct positions of eyes are likely to be gathered to one point because faces and eyes are detected from the facial records classified by face and eye classifiers while reducing images. Therefore, an average excluding error range is calculated to minimize erroneous detection.
While the invention has been shown and described with respect to the embodiment, it will be understood by those skilled in the art that various changes and modifications may be made without departing from the scope of the invention as defined in the following claims.
Number | Date | Country | Kind |
---|---|---|---|
10-2008-0131783 | Dec 2008 | KR | national |