During fall semester, I took a great class called Computational Vision. Taught by Dr. Thomas Serre, the class focused on biological models of vision. Topics covered included color, simple and complex cell models, motion processing, stereo vision, and object recognition.
For my final project, I tried to estimate age from facial images. I used the FG-NET aging database, which contains roughly 1,002 color and grayscale facial images. My method is roughly outlined below
- Locate facial features
- Find center of pupils, mouth, nose, outline of jaw
- Compute facial feature ratios
- Wrinkle analysis
- Mask wrinkle-prone areas
- Filter areas with bank of Gabors
- Compute texton dictionary
- Compute bag of textons representation
- Run classification/regression
Binary classification between infant and non-infant classes resulted in 97.2% classification accuracy and 97.8% area under the ROC curve, which is pretty good. To see more results, including regression results, read the paper here.
Here’s a gallery of some of the assignments for the class. You can actually see most of the course materials at https://canvas.brown.edu/courses/350884.