Comparative Analysis of CNN Based Face Detection

Chenxing Ma

Abstract: The face detection task has been studied in depth. There are may efficient face detectors which utilize specialized designs in different aspects for the detection task for faces, making the detection algorithms and models more and more complex. As a result, the computational and time cost becomes higher. In recent years, many studies are carried out aiming at reducing the algorithm and model complexity. These simpler face detectors make the detection faster while ensuring detection accuracy. In this paper, we select three different face detection models that simplify the face detection algorithms or model structures based on common CNN networks and YOLO structures, they are, YOLO5Face, DSFD and TinaFace. We first analyze the algorithm and model structure of the selected face detectors and then test them on several datasets to evaluate the generalization ability of the models. The experiment result show that the selected face detectors can efficiently complete the face detection task while YOLO5Face has the best performance on the datasets.

Keywords: Face detection, CNN, YOLO, datasets