A novel hybrid ensemble convolutional neural network for face recognition by optimizing hyperparameters

Abstract

A fully fledged face recognition system consists of face detection, face alignment, and face recognition. Facial recognition has been challenging due to various unconstrained factors such as pose variation, illumination, aging, partial occlusion, low resolution, etc. The traditional approaches to face recognition have some limitations in an unconstrained environment. Therefore, the task of face recognition is improved using various deep learning architectures. Though the contemporary deep learning techniques for face recognition systems improved overall efficiency, a resilient and efficacious system is still required. Therefore, we proposed a hybrid ensemble convolutional neural network (HE-CNN) framework using ensemble transfer learning from the modified pre-trained models for face recognition. The concept of progressive training is used for training the model that significantly enhanced the recognition accuracy. The proposed modifications in the classification layers and training process generated best-in-class results and improved the recognition accuracy. Further, the suggested model is evaluated using a self-created criminal dataset to demonstrate the use of facial recognition in real-time. The suggested HE-CNN model obtained an accuracy of 99.35, 91.58, and 95% on labeled faces in the wild (LFW), cross pose LFW, and self-created datasets, respectively.

Keywords

• ensemble learning
• face detection
• face recognition
• hyperparameters tuning
• transfer learning