CSE455 Project 4

Albert Chiu

Average face along with 10 eigenfaces

As we can see, for facial recognition there are diminishing returns for increasing the number of faces in the database. While the number of faces properly recognized does increase as we add more faces, an excess of faces will lead to additional errors in checking if something not in the user base is a face. There is no clear solution as to which number of faces is ideal, but for this example somewhere between 7 and 21 is likely where the ideal number is. What the primary purpose of the software is will determine where on this scale the ideal number is.

Images 2 and smiling 11

Here we can see the mix up between 11 and 2. I doubt a human would make this mistake, but the error does seem somewhat reasonable. The 11^th image actually is the next in the list of possibilities, so this wasn’t a huge mistake.

Images 16 and smiling 14

Here we can see the mix up between 14 and 16. This mix-up seems a bit more unreasonable, as the faces are quite a bit different. 14, however, doesn’t appear in the list until the 11^th entry, meaning the computer was fairly lost about this face.

Cropping and finding faces

Cropped elf

Crop and mark run with scale between .15 and .35 with step .01

It was surprising that the program was unable to properly crop my own face (am I that abnormal looking?). There is a clear lighting gradient across the face that may have caused errors in the program. In addition, my face is somewhat longer than most people’s, which may have helped lead to the error.

Image recognition of smiling and nonsmiling students. Ran with min scale .35 to max scale .45 step .01

Classroom face recognition. Run with scale between 1 and 2 with step .1

Only 6 out of the 12 students were identified (50%). Errors here may well be due to different lighting along with a diverse number of facial expressions. In addition, larger images lead to more possibilities of areas that are not actually faces to be taken as false positives.

Run with scale .25 to .45

Ignoring the bottom left image (since the scale is likely too small for it), this image recognized 3 of the 5 images, and caught a part of a 4^th. This image is particularly hard to identify faces in due to the different facial expressions on each, the face facing different directions, and different lighting.

Verification

Using a batch script, I ran every image in non-smiling against every image in smiling and gathered all the MSE information. Below, you’ll see the number of false positives that occur for every additional correct answer.

As you can see, getting every single one requires a very large number of false negatives. The best MSE threshold seems to get either 17 or 21 of the 24 correctly verified. These thresholds are 45271.1 and 79021.7 respectively. They result in 33 and 104 false positives while failing to correctly verify 7 and 3 of the faces.