A simple cascaded convolutional neural network for single face detection implemented in python using Keras. Due to the computationally heavy nature of deep networks a shallow initial net, 24Net, is implemented for fast discrimination of the sub-images generated by a sliding window together with an image pyramid. This allows for a small portion of the input data to be passed to the deeper 48Net for final detection. The architecture enables computationally efficient detections with an accuracy of 88%.
To run the code you need Python 2 and Keras installed and a dataset. We used FDDB.
Code entry point is main.py. Read project_report.pdf for a thorough review of the method.