Face Recognition - Algorithmic Approach For Large Datasets And 3D Based Point Clouds

摘要

This work proposes solutions for two different scenarios in face recognition and verification. The first scenario involves large scale unconstrained unsupervised face recognition. The proposed system for this scenario is a complete face recognition framework. The proposed system first studies the performance of unsupervised face recognition for frontalized captured faces in the wild under the effect of a single image super-resolution algorithm. The system also introduces new high dimensional features based on LBP and SURF that perform better than the state-of-the-art features for unconstrained unsupervised face recognition. To solve the large scale recognition process, a new algorithm has been designed to manipulate face images in the dataset. This new algorithm represents all training face images as a fully connected graph. The algorithm then divides the fully connected graph into simpler sub-graphs to enhance the overall recognition rate. The sub-graphs are generated dynamically, and a comparison between different sub-graph selection techniques including minimizing edge weight sums, random selection, and maximizing sum of edge weights inside the sub-graph is provided. Results show that the optimized hierarchical dynamic technique developed with sub-graphs selection increases the recognition rate in large benchmark image dataset by more than 40% for rank 1 recognition rate compared to the original single large graph method. The approach developed in this research is tested on different datasets, especially if the number of images per person in the training data is low. Furthermore, in order to improve rank 1 recognition rates and to reduce the computation time of the recognition process, a new technique that combines the hierarchical face recognition algorithm and a deep learning neural network using Siamese structure for face verification is proposed. The second part of this work addresses the usage of neural generative models for 3D faces with an application in face recognition when 3D datasets are utilized separately without the existence of texture information scenarios. An improved technique is developed to construct new representations for point clouds containing 3D information. The technique employs a regression neural network model trained using Levenberg-Marquardt (LM) algorithm. One of the advantages of this new representation is the significant reduction in storage space required for point clouds due to the utilization of a regression model for depth map regeneration. Moreover, the trained neural models can be used to generate a super-resolution version of the original 3D point clouds. The proposed regression representation is also used with a deep Siamese neural system to implement a complete depth-based neural face recognition and verification framework. The results indicate that the proposed system provides highly accurate and efficient face recognition results with 3D information only without texture information.