Face detection and recognition

Home page | Curriculum Vitae | Acid | Face detection and recognition | Papers | Wonder computer | DarkCut | Dark | Environment | Tips | Links |

My PhD work is on face detection and recognition. I am currently working on a complete face detection and recognition, in color images with a complex background. I have already good results for my algorithms. This web page will be updated as the project will grow.


Project status

A modified skin segmentation based on Terrillon et al. work was successfully obtained. I now can retrieve from an image with complex background skin areas. See the text bellow for problems.
The face detector is now completed using Viola Jones based method. Very good results are obtained, since the color segmentation "cleans" almost all the complex background, making the system more robust.
Next task: Finding face features to obtain only faces from the image.


Important facts

  • If you are a researcher in the same field, I want to compare my results (on skin segmentation) to others projects. For what I know, there are not any databases for skin segmentation or even face detection on color images. Do you know any ?
  • Are you doing almost the same as I do ? Lets compare your work, and results !

  • News

    [24-Ago-2004] The face detector performs very well using a Viola Jones based method. Not too fast, but accurate.
    [22-May-2004] The segmentation is even faster, a new paper was published about this subject.
    [12-Nov-2003] The segmentation (ver 1.3) is stable and robust. Now working on face detection.
    [30-Set-2003] The segmentation program is available for download since version 1.0. Give the necessary credits when comparing !
    [30-Ago-2003] Added a new image, present in another skin detection method. This algorithm alse performs good to the test image (dark skin men playing golf).
    [28-Ago-2003] The skin segmentation algorithm is now completed. I am now working on face detection based on the skin segmentation.
    [17-Ago-2003] The skin segmentation algorithm is at its final phase. I am accepting images for processing and testing. Use the mail in the footer.

    Download

    You can test my work. Download the colorseg program from below. This do not include source files, but the segmentation is performed. Enjoy !

    File


    colorseg1.3.tar.gz
    fdr0.8.tar.gz

    Description


    ColorSeg program (ver 1.3)
    Face Detection (ver 0.8)

    File size


    2279304
    1965632

    Last update


    12 Dec 2003 15:46:55
    24 Aug 2004 15:26:30

    Skin segmentation

    What purpose?

    The need for better human computer interaction is now a fact and will probably be in the next few years. More friendly and effective methods regarding human activity are in constant development to free users from manual interaction. Gesture recognition, robot interaction, multimedia, face detection and recognition, hand detection, teleconference and many other applications are, or can be based on skin detection to restrict the complexity of further processing. Size, quality and cost of image acquisition are also very important to the rapid growing needs for such systems and commercial applications are now available.

    Method?

    I use a color segmentation method similar to the one presented by Terrillon in the paper Automatic Detection of Human Faces in Natural Scene Images by Use of a Skin Color Model and of Invariant Moments, 1998. Some modifications were made to the scaling method and 'better' (no database for comparison) results are now achieved, since the skin of Caucasian, Asian, African and descents are segmented.

    Problems

    This method if very robust, but also vulnerable when objects are present with a color similar to skin, since we base the segmentation in color. This is the main disadvantage.

    Points in favor

  • Not restricted to orientation, size or grouping candidates
  • Can handle a complex background
  • Cover all skin colors
  • Results

    Some images are present that shows the system abilities:
    Conditions:
    Very complex picture with different sizes groups of skin. Showing the original image, the median image and the group resulting analysis. Only one group is not correctly classified since it has a very similar color to skin. The grouping algorithm also connects two separated elements (forehead from the rest of the face) that compose the man with glasses.
    Image size: 800x600.
    Input device: Sony P51.
    Comments: Night picture with flash, indoor.
    Conditions:
    Complex picture with frontal face and high light variation. Showing the original image, the median image and the group resulting analysis. The effect of light difficult the correct extraction of the face.
    Image size: 640x480.
    Input device: Philips ToUCam.
    Comments: Daylight, indoor.
    Conditions:
    Simple background with different color skin arms and hands. Showing the original image and the median image. The effect of light is observed in the bigger hand and on the arm present on the right top corner.
    Image size: 1600x1200.
    Input device: Sony P51.
    Comments: Night with artificial light and flash, indoor.
    Conditions:
    Very simple background with a dark skin color golf player. Taken from the web in another skin segmentation method web page.
    Image size: 209x222.
    Input device: Unknown.
    Comments: Outdoor picture.
    Filipe António Gonçalves Tomaz
    This page is hosted in the University of Algarve
    Last update: Tue, 24 Aug 2004 15:30:31 GMT