Reduced Residual Nets (Red-Nets): Low Powered Adversarial Outlier Detectors

The evolution of information science has seen an immense growth in multimedia data, specially in the case of CCTV live stream capture. The tremendously large volumes of multimedia data give rise to a particularly challenging problem called the outlier events of interest detection. In the wake of growing school shootings in the United States, there needs to be a rethinking of our security strategies regarding the safety of children at school utilizing multimedia data mining research. This paper proposes a novel method to identify faces of interest using live stream CCTV data. By integrating the adversarial information, the proposed framework can help imbalance facial recognition and enhance rare class mining even with trivial scores from the minority class. Experimental results on the Faces in the Wile (FIW) dataset demonstrate the effectiveness of the proposed framework with promising performance. The proposed method was implemented on a low powered Nvidia TX2 for real-time face recognition. The proposed framework was benchmarked against several existing state-of-the-art methods for accuracy, computational complexity, and real-time power measurement. The proposed method performs very well under the power and complexity constraints.