3D Super Resolution using Auto Encoders for Face Recognition

In order to perform 3D face recognition with good performance high quality images are required, however high resolution 3D cameras such are the Minolta vivid 910 are expensive. Cheaper and more widely available camaras such as the Microsoft Kinect and the Intel Realsense have lower quality. This research will present a method to apply super resolution and improve the quality of these low resolution images using auto encoders. Existing methods are all based on interpolation, which are unable to reproduce high frequency components of the image, with auto encoders it is assumed that this is possible and previous research in 2D super resolution with auto encoders has shown that this is indeed the case. This research will look into different auto encoder architectures training them down sampled and noisy versions of high quality databases. The best architecture will be chosen and will be compared to an implementation of interpolation by performing face recognition on the reconstructions and doing error analysis using the original high quality images. To further test both methods lower quality images from low quality database will also be reconstructed and compared. The results have shown that the autoencoder is indeed better in this case.