Learning-based Face Detection

1. Feature optimization. Two training sets, one of face patterns (images) and the other of non-face patterns, are constructed.
   
   A canonical training face sample

   The trainig patterns are preprocessed to normalize the image intensity values. They are then group into several clusters of faces and non-faces.

   
   Centroids of six face (left) and six non-face clusters.

   Each cluster is approximated by a Gaussian distribution. Principal components of the distribution is extracted and used to compute the distribution more accurately while components of smallest eigenvalues are discarded. After that, LDA is applied to maximize separability between clusters.

2. Neural learning. Given the optimized feature sets, a complex mapping function for the classification is learned from the feature sets to produce the minimum error rate. A backprop learning procedure is used.

3. Neural classification. This is to classify the pattern in a given image window into face or non-face. Multiscale search strategy to detect faces with different scales. Multiple Network results are combined by arbitration.
   
   Multiple neural networks with arbitration for classification.

The following show some face detection results: