3D morphable model fitting for low-resolution facial images

This paper proposes a new algorithm for fitting a 3D morphable face model on low-resolution (LR) facial images. We analyse the criterion commonly used by the main fitting algorithms and by comparing with an image formation model, show that this criterion is only valid if the resolution of the input image is high. We then derive an imaging model to describe the process of LR image formation given the 3D model. Finally, we use this imaging model to improve the fitting criterion. Experimental results show that our algorithm significantly improves fitting results on LR images and yields similar parameters to those that would have been obtained if the input image had a higher resolution. We also show that our algorithm can be used for face recognition in low-resolutions where the conventional fitting algorithms fail.