Facial_Emotional_Classifier
We are using the Cohn-Kanade Facial Expression extended database (http://www.pitt.edu/~emotion/ck-spread.htm)
We cannot distribute it so you will have to request it yourself (from the link above)
DatasetOrganiser.py contains the code to organise the dataset.
Image pre-processing options like mean filter, gaussian filter, median filter and bilateral filtering are available. They can be turned on in the Config.py file.
It's best if the images are all of the same size and have only a face on them (no clutter). Thus, We find the face on each image, convert to grayscale, crop it, apply some smoothening techniques and save the image to the dataset. We a HAAR filter from OpenCV to automate face finding.
To build an image representation, we use Bag of Words (BoW) and Fisher Vector (FV) approaches. Both BoW and FV featurizations rely on a beforehand computation of local descriptors of images. We choose to consider the very popular SIFT (Scale-Invariant Feature Transform) descriptors. Very interesting properties of these descriptors with respect to our problem are their invariance to affine transformations, and robustness to changes in illumination, noise, and small changes in view point.
For those who are more comfortable learning through video, check out https://www.youtube.com/watch?v=NPcMS49V5hg For people who like reading, see this http://docs.opencv.org/3.1.0/da/df5/tutorial_py_sift_intro.html#gsc.tab=0
Now with the idea to get more distinct and compact descriptors (Dimensionality Reduction), we perform a Principal Component Analysis (PCA) on the SIFT descriptors.To get more information on PCA, check out https://en.wikipedia.org/wiki/Principal_component_analysis
Bag of Visual Vectors approach is now applied using kMeans clustering to build the image representation and each image is now a histogram of these clusters
** Current Model Accuracy**
