Using CNN model to recognize the facial keypoints over different faces in the image
The CNN structure has a unique structure that contains five convolution layers with composed Pulling Convolution Layers for compasing features.
- Normalizing the image to binary image and keypoints in range -1~1
- Rescaling the images to 224x224
- Random crop
- Random rotation about 15 degres
Using Adam optimizer and Smooth L1 Loss as a regression loss function.