Resolution-aware Ensemble of Pose and Illumination-Invariant Feature Descriptors for Face Identification in Unconstrained Videos

Fagbola, Temitayo Matthew; Thakur, Colin Surendra

Abstract

The recurring global security insurgence has posed new opportunities for massive deployment of Video Surveillance Systems (VSS). However, videos captured by such systems suffer characterized Low Resolution (LR), varying illumination and pose challenges of subjects present in the videos. Consequently to manage these limitations, most existing Feature Extraction Techniques (FET) lack support for the limitations inherent in most VSS videos, which accounts for the high computational overhead and low accuracy of most video-based Face Recognition Systems. In this paper, Iterative Back Projection-Maximum A Posteriori (IBP-MAP) resolution reconstruction technique and an ensemble of local and global feature descriptors based on Linear Discriminant Analysis (LDA), Local Binary Pattern (LBP) and Gabor Wavelet Transform (GWT) were used to realize an improved pose and illumination invariant FET suitable for LR videos. 2900 LR frames were obtained from YouTube Celebrities corpus and a Locally Acquired Video Dataset (LAViD). These frames in the range of 30 pixels to 65 pixels were reconstructed using IBP-MAP. LDA-LBP-GWT ensemble was developed by fusing the facial features of Linear LDA, LBP and GWT into a LDA-LBP-GWT Single Feature Set (SFS). The SFS was dimensionally reduced using particle swarm optimization algorithm. The LR and the reconstructed frames were used as testing sets while locally acquired pose-oriented mugshots constituted the training set. Features of each frame in the testing sets were compared with those in the training set for recognition using Euclidean distance. The developed techniques were implemented in MATLAB 2019. The performance of the developed LDA-LBP-GWT ensemble was compared with the baseline techniques by using False Acceptance (FA), Recognition Accuracy (RA), Recognition Time (RT) and False Rejection (FR) as evaluation metrics. Results obtained indicate that the developed LDA-LBP-GWT ensemble serves as improvement over the baseline techniques in terms of FA, RA and RT.

Citation

Fagbola, T. M., & Thakur, C. S. (2019). Resolution-aware Ensemble of Pose and Illumination-Invariant Feature Descriptors for Face Identification in Unconstrained Videos. International Journal of Engineering Research & Technology, 12(12), 3114-3126

• Journal Article Type: Article
• Publication Date: Jan 1, 2019
• Deposit Date: Jan 28, 2024
• Journal: International Journal of Engineering Research and Technology
• Peer Reviewed: Peer Reviewed
• Volume: 12
• Issue: 12
• Pages: 3114-3126
• Public URL: https://hull-repository.worktribe.com/output/4161539
• Publisher URL: http://www.irphouse.com/ijert19/ijertv12n12_142.pdf