Face Recognition and Detection

by Xiao Li

Face recognition

Average face

Top 10 eigenfaces

Plot: Correctness vs. # of eigenfaces

Analysis:

1. The recognition correctness goes up quickly as the number of eigenfaces increases from 1 t o 5. It fluctuates between and 5 and 19 and stay flat after 21. More number of eigenfaces means more computation. In fact, the eigenfaces with low eigenvalues are trivial in the discriminality and hence can be ingored. 11 eigenfaces is a good choice in this sense.

2. An example of false recognition:

In this case, using 11 eigenfaces, Zhou is recognized as Ko. It seems reasonable, since Ko's unsmiling face has some characteristics of Zhou's smiling face.

Face Detection

is cropped as

with parameters min_scale = 0.45, max_scale 0.55, step = 0.01

is cropped as

with parameters min_scale = 0.18, max_scale 0.24, step = 0.01

with parameters min_scale = 0.95, max_scale 1.05, step = 0.01

with parameters min_scale = 0.9, max_scale 0.10, step = 0.01

This is a hard case, since there are high texture areas all varound. Our faces were sided and the lighting was not good. It is reasonable for the two erros, since one was wearing a big glasses and my face was shadowed by the hat. The one below looks better :)