DeepView: A Wireless Dynamic Facial Recognition System with Data Logging

Authors: Blaze R.Perater, Benjamin R. Sanglitan, and Cristina P. Dadula

Abstract

Conventional facial recognition techniques are nonversatile to changes in pose expression because they utilize static algorithms. This paper proposed a dynamic wireless facial recognition system with data logging capabilities using CNN. Face recognition methodology was divided into two stages: face detection and face recognition. For face detection, a Histogram of Oriented Gradients (HOG)-based technique was used in conjunction with Face Alignment through Affine Transformation for input image pre-processing. The facial recognition stage utilized an OpenFace implementation for the neural network, modified clustering for grouping identities and Density-Based Spatial Clustering of Applications with Noise (DBSCAN) for removing outliers. The accuracy was calculated at 87.03% with an average processing time of 13.7 ms at 10 fps frame rate. Images are sorted in archives of the data logger by time, date, camera number and picture of encounter for each distinct identity.
In addition, face searching enables the user to upload and external photo and search the database for a matching identity. The system has been successfully implemented in a real world scenario.