LiteFace: A Light-Weight Multi-person Face Detection Model

Deep neural network based face detection models have made remarkable progress in recent years. One-stage detectors have been widely used due to their balance between speed and accuracy. In this study, a light-weight face detection model, LiteFace, that uses YOLOv5 model as a baseline is proposed. This model detects and counts the faces in a given image with the addition of the focal loss function to mitigate the problem of imbalanced training data. Wasserstein loss is used to focus on small-scale faces, while Intersection over Union (IoU) loss is used for large-scale and medium-scale face detection. In order to achieve receptive field enhancement of the upper pyramid layers of the network, an Inception module is introduced at the top of the bottom-up pathway. A Convolutional Block Attention Model is added before the detection heads to help the network focus on regions of interest. This network is trained on a subset of the training images from the WIDER FACE dataset and achieves an improved training time with an accuracy of 94.1% on the easy subset of WIDER FACE.