This project uses images collected from image.google.com to train an emotion classifier.
It's still under development, and the test accuracy is 52%.
-
Data collection: search & batch download on image.google.com.
-
Data processing(process_original_images.py): convert profile images to face images using OpenCV.
-
Network building: currently I use an extended LeNet-5.
-
Training: AdamOptimization or AMSGrad.
-
Predict: classify the input image that contains at least one human face.
# Will train the network first if a pretrained one doesn't exist.
$ python emotion_recognition.py katy_perry.jpg