A facial expression classifier created using OpenCV and large CNN, with numerous techniques used to increase the accuracy of the model.
The dataset was initially retrieved from Kaggle, with a direct link to the files also being provided:
https://drive.google.com/drive/folders/1bCFKTkmItIZl73YNUq5QXppLpIdmTgUv
Methods used to Increase Accuracy of the Model:
-Data Generator, used on training set: rotations, shearing, horizontal and vertical shifts, zooms, horizontal flip, shuffling dataset
-He normal initialization for ReLU layers and Xavier (glorot normal) initialization for softmax layer
-Batch Normalization and image padding
-Adam optimizer (lr=0.001)
-ModelCheckpoint, Learning Rate Reduction, Early Stopping
Result:
First Round of Training: Final Accuracy = 47.29%, Validation Accuracy = 38.38%
Second Round of Training: Final Accuracy = 58.33%, Validation Accuracy = 54.99%
For the OpenCV detection code, the Haar Cascade algorithm was used to detect faces from the webcam's input. Originally proposed in the paper titled "Rapid Object Detection using a Boosted Cascade of Simple Features", the Haar Cascade algorithm returns the coordinates of the corners of a rectangle drawn around the face of the person. This allows for the classification of a cropped image of the person's face, rather than the whole input image, which improves accuracy.
Link to the original paper and code:
https://www.cs.cmu.edu/~efros/courses/LBMV07/Papers/viola-cvpr-01.pdf
https://github.com/opencv/opencv/blob/master/data/haarcascades/haarcascade_frontalface_default.xml