250 likes | 395 Views
RGB Level Face Detection. Jian Zhang Miao He Jing Chen May.27th,2002. How to find faces?. Mission Analysis. lots of faces in the images, time-consuming to search the face candidates, some faces got overlapped
E N D
RGB Level Face Detection Jian Zhang Miao He Jing Chen May.27th,2002
Mission Analysis • lots of faces in the images, time-consuming to search the face candidates, some faces got overlapped • sidelight instead of diffused light. Influence from shade, complexion, rotation of face, and the inhomogeneity of background • 1280*960 high resolution image, slows down the detection speed and more noise
2-step Algorithm • 1. Skin Toner Masks • 2. Support Vector Machine (SVM)
Skin Toner Masks (1) • Mask based on hue in HSV space
Skin Toner Masks (2) • Mask based on RGB statistics
Skin Toner Masks (3) • Mask for red color
Skin Toner Masks (4) • Mask to remove the ground
Skin Toner Masks (5) • Remove small regions
SVM Theory(1) • Given the training sample {(xi,di)}, i=1…L, find the Lagrange multipliers {αi}, i=1…L, that maximize the objective function • Subject to the constraints 1) 2) for i=1,2…L Where C is a user-specified positive parameter.
SVM Theory(2) • The discriminate is
Pre-processing • In the tradeoff between speed and accuracy, we choose 12-by-12 and 4 gray scale samples. • Eliminating the effect of side light. • Histogram equalization
Training a SVM • a large date set, about 1000 face (different scales and positions) and more than 7000 non-face samples. • takes more than 30 hours on the ISL lab computer to get one set of training result. • From now on,SVM shows its advantage
Our improvement • Add rotated face into face data set. Thus we can detect these special faces • faces concentrate against non-faces B is an observation constant. Thusenhance the detection accuracy.
Testing Results (1) • Now we need merge the detection points to give the final decision. Model the three-scale detection as a diversity situation like in wireless communication channels.
Testing Results (2) • maximum-ratio diversity • perform dilation to binary image
Acknowledgement • The authors want to express out thankfulness to Prof. Girod’s excellent instruction. We all learn a lot from this interesting course. • We would also like to thank our teaching assistant, Chuo-ling Chang, for his patience and help.