Improving Facial Recognition of FaceNet in a small dataset using DeepFakes

Abstract

Facial recognition is affected by many factors such as low-resolution images, availability of datasets, illumination, pose invariance, ageing, expression, etc. With the increasing availability of powerful GPUs, we are using massive datasets to facilitate better accuracy. Among different datasets available to perform training of any facial recognition algorithm, very few of them can be a run on a low configuration system. Still, the same can’t be used to create satisfactory FaceSwap results because of anomalies in them. A small dataset of 20 identities has been created on which the results of this paper are observed. This paper introduces the usage of DeepFakes algorithm to improve the performance of FaceNet with SqueezeNet architecture and softmax loss function. It is expected that more data leads to better performance. Then the FaceSwap variation of DeepFakes is used to swap identities and create fake images for a given identity. Then, FaceNet is used to identify faces on the newly formed dataset 250RF using fake images and the original dataset 200R. This method achieves satisfactory results on training and testing accuracy in comparison, thereby creating prospects on such a method.