Lightweight CPU-based Face detection with Efficient Feature Selector in Real-time Applications

Face detection is the initial process of a face analysis system. It is thus crucial for intelligent surveillance systems that require an smart vision method to operate in real-time scenarios. The face detection capability will increase the practical value when implemented on a low-cost device, such as a CPU. Several lightweight face detection architectures exist; however, they require many parameters and computations. In this paper, a rapid face detector is proposed that localizes the facial area for implementation of video surveillance and Human-robot interaction systems on cheap devices. The detector includes a feature selector, which distinguishes the essential element using the positional relationship between features. In addition, an attention module is used to improve the feature quality by capturing global features on each mutually exclusive detection layer branch. Experimental results show that the proposed detector achieves a competitive performance of state-of-the-art lightweight face detection using only a few parameters and operations. Moreover, it has a fast data processing speed that reaches 122 FPS on a Core i5 CPU.