| |
| |
| Face Recognition |
|
| |
| Average Face |
 |
| |
| Top Ten Eigenfaces |
 |
| |
| Face Recognition Results |
 |
| The above graph shows the number of faces correctly recognized (out of 27), while varying the number of eigenfaces used. As you can see the number of correctly identified faces quickly grows until around 11 eigenfaces are used. After this the accuracly levels off staying relatively constant. This is most likely caused by the eigenfaces with smaller eigenvalues not being representative of facial features, and so adding them provides no benefit to the recognition. |
| |
The incorrectly recognized faces were also consistently recognized as a different person no matter how many eigenfaces were used. Below shows a few of the most common misrecognized faces.
| Original Face | Recognized As | Difference | Correct Face in Results |
 |
 |
 |
#2 |
 |
|
 |
#3 |
The difference of the first false positive shows that there is very little that is different between the faces and it is easy to see why the recognized face was incorrect. The second shows a lot more variation around the eyes and mouth but the eigenfaces used seem to be unable to handle this variation.
|
| |
| |
| Finding Faces |
|
| |
elf.tga
  |
Cropped:
 |
| My program seemed to have found a decent match for the face even though it was not included in the training sample. Suprisingly it was also able to successfully find the face of the baby. |
| |
Self Photo
 |
Cropped:
 |
| Again my program seems to find a good match for my own face when searching 0.45-0.55 with step 0.01. |
| |
IMG_0031.tga
 |
| The three faces in this photo seem to be correctly detected when using the 0.45-0.55 with step 0.01. |
| |
Random Group Photo
 |
| With this image my program seemed to have quite a bit of trouble finding the faces in the image only successfully finding 1 when using a search window of 0.85-1.0 with step 0.01. Two of the "found" faces included parts of the arms. It also seemed to have trouble with the textured areas finding part of the tree and grass areas. There are probably several reasons why the faces were not detected. First of all the training set is rather small and not representative of a large variety of faces. Second my program seems to have problems with regions of small variation despite trying to correct for this, which caused some problems. |