Discarding Non-Faces before the Recognition Stage

At the recognition stage, we have a ``DFFS'' value, the 60-element key, and the residue value of a fully normalized mug-shot image. First, if the ``DFFS'' value is above a threshold (specified by the user), the mug-shot image does not contain a true face. In other words, it is quite far from the cloud of faces in our training set and should probably be discarded from the recognition algorithm. Recall the face localization procedure presented in Chapter 3. It is possible that the face localization incorrectly converges to an image that is not a face. The detection process described in Chapter 3 only dealt with blobs and limbs. This simplistic description of image data allows us to detect faces in many poses, although this flexibility might also permit a non-face to be falsely detected by the algorithm. A soccer-ball, for example, might look like a face with two eyes, a mouth and a nose (from a simple blob and limb description of the image). Since the face-detection scheme described on Chapter 3 is based exclusively on a simple blob and limb description of the image, the recognition stage might obtain a mug-shot of a soccer ball or other non-face object. However, in the recognition stage, we utilize the statistically based ``distance to face-space'' measurement to ultimately reject non-face mug-shots from the localization procedure. Thus, the mug-shots generated by the face-localization process can be filtered with a threshold on their ``DFFS'' value before recognition is attempted. Equation  illustrates the use of a threshold on the ``DFFS'' value to prevent recognition attempts on non-faces. Typically, the value used for DFFS_threshold is roughly 3000.

 

DFFS < DFFS_threshold (4.43)