Deep-learning-based Face Detection And Segmentation Using Iterative Bounding-box Regression

With the rapid development of artificial intelligence(AI),the machines are becoming more intelligent.In the field of artificial intelligence,the human face is undoubtedly an important research object.The first step of face research is face detection,which is to detect the face and get the exact location from the image.Our intelligent machines will serve human beings better if they can accurately detect all faces in images or videos.Multi-view face detection in open environments is a challenging task due to the diverse variations of face appearances and occlusion.In fact,the task of face detection includes classification and localization,the localization accuracy is one of the key factors.Accurate location can preserve useful information while reducing redundant information,which is beneficial for subsequent recognition algorithm.However,many of the existing methods do not pay enough attention to localization.Some of the current methods have applied localization techniques,but they have not fully realized its potential and realized more accurate localization.Face segmentation is of great significance in the application scene such as background stripping.Face segmentation is introduced to our face detection task as an auxiliary task,which provide segmentation function,and improve the localization and overall performance by utilizing the internal connection between the two tasks.In this paper,we focus on improving face detection perfromance by enhancing the localization ability.This paper proposes a localization technique,called iterative bounding-box regression,to approach the potential faces.Further more,an pyramid pooling based multi-task segmentation network is designed to get the meticulous depiction of face in pixel-level.From the whole point of view,this paper proposes a deep cascaded detection method to detect faces.Firstly,we use a full-convolutional network to detect faces,and then we iteratively exploit the bounding-box regression to approach the potential faces.Besides,we exploit segmentation network to further boost up overall performance by re-classifing and segmenting.Extensive experiments demonstrate the better localization accuray of our algorithm by comparing it with several popular face detection algorithms on the widely used AFW dataset and FDDB dataset,and it also prove that the improvement of localization abilit y can boost up the overall performance of face detection.