An Efficient Face Detector on a CPU Using Dual-Camera Sensors for Intelligent Surveillance Systems

Intelligent surveillance systems require face detection to identify human facial areas. This system should be able to utilize dual-camera sensors (color/IR) for working every time. Additionally, practical application demands a detector to be operated in real-time on a low-cost device or CPU. The deep Convolutional Neural Network (DCNN) technique has successfully used a robust facial extractor, but it requires a high amount of computation for high-resolution input. On the other hand, the light architecture generates a large number of false positives as a fast detector. This feature extractor pays less attention to specific facial features and often ignores global and local relationships between elements. This paper proposes an efficient face detector to accurately localize faces using light architecture. The one-stage detector consists of an efficient backbone to rapidly extract features and a four-level detection layer to predict variations in facial scales. To improve the non-robust feature extractor, it implements an enhancement module to enhance specific facial features at each level without significantly increasing the parameters. The proposed detector uses knowledge from the WIDER FACE dataset to train the model with a gradual learning rate. The experiment results show the effectiveness of the detector in outperforming CPU-based detectors on benchmark datasets. It also runs in real-time at 27 frames per second on a CPU using the RGB and infrared cameras for Full High Definition (Full HD) resolution, faster than other published detectors.