Rotation Invariant
Neural-Network Based
Face Detection
Overview
Multiple Neural Networks
Router Networks
Detector Networks
Overview of how the algorithm works
Input and output of the
router network
Rotation Network:
Outputs are generated as weighted vectors
Average of the weighted vectors is interpreted as an angle
1048 training images labeled by face, eyes, tip of the nose,
corners and centers of the mouth
Each training face is rotated 15 times in a circle
Rotation Neural Net
Description
400 layers on the input layer (20X20)
Hidden layer of 15 units, output layer
of 36 units.
Hyperbolic tangent activation function
Standard error back propigation
Detector Network
Identical to the routing network.
Trained by positive (contains faces)
and negative images (does not contain
faces).
Weights are initially random for the
first training iteration.
Training on non-face images, add false
positives to the non-image
Adding False Positives to the training set
as negative images
Arbitration Scheme
Detection of Different Faces at
different angles in the same image
Detections are placed in 4 dimensional
space - x,y,angle, pyramid level,
quantized in 10 degree increments.
Two independently trained networks
are ANDed to improve the success rate.
Empirical Results:
130 images, 511 faces
Sample Images
Conclusions
Represents ways of integration
multiple neural nets
Speed of implementation
Face Detection VS Facial Recognition