Face Recognition in Unconstrained Environment Using Deep Learning

Abstract

Face recognition as a biometric is used in many applications as it does not require much cooperation from the user. The existing face recognition techniques give good results in constrained environment but their performance in unconstrained environment degrades as the face images captured vary in pose, illumination and resolution. The present study tries to develop a robust face recognition model in unconstrained environment using deep learning. To achieve this, the first step is to create a novel face dataset by collecting images of celebrities from Internet and images captured from real-world surveillance videos. Deep neural network (DNN) available in OpenCV is used for face detection. Second, Deep convolution neural network model with eight convolution layers was developed and trained on this dataset. The proposed model achieved an accuracy of 99.75%, whereas fine-tuned VGGFace achieved 99.6% and Alexnet trained from scratch on the dataset achieved 98.6% on the test data. The proposed model has 5,671,824 parameters as compared to VGGFace fine-tuned with 40,487,824 and Alexnet with 28,159,832 parameters. The size of the proposed model is 68 MB in comparison with VGGFace and Alexnet having 293 and 338 MB, respectively. The impact of batch size hyperparameter on the model was also studied, followed by comparison between RMSProp and Adam optimizers. The results demonstrate that with the increase in batch size, the performance improves and Adam optimizer outperforms RMSProp.