Anchor-free face detection in the wild with an application to facial expression recognition

Facial expression recognition (FER) has been received more attention in the field of affective computing. FER starts from the basic tasks of bounding box detection and landmark localization. For the sake of simplicity, we call bounding box detection and landmark localization as face detection. In this paper, we aim at proposing a new face detection method for FER system to improve the detect precision. More specifically, we present a new face localization based on a full convolution one-stage object detection algorithm, namely FCOSFace. It utilizes the advantages of Feature Pyramid Network and center-ness loss, which extract the shallow and high-level semantic features and suppress the influence of negative samples. Furthermore, based on FCOSFace, we used the landmark detection with DLib on the detected face. In order to evaluate the efficiency of the proposed system, experiments on both face localization and FER databases have been conducted. Experimental results demonstrate our proposed method has well-performance on face detection. Lastly, we extend our proposed method in facial expression recognition, especially in two challenging emotion databases including FER2013 and SFEW-SPI. Results on facial expression databases see the our designed system has promising results on facial expression recognition.