Face Detection And Face Alignment Algorithm Using Multi-Task Learning

Face detection and alignment is one of the most active fields of computer vision.Face detection attempts to localize faces in the image.Face Alignment further aims to localize facial semantic landmarks from a given face area.The detection speed and accuracy of these two tasks seriously affects the performance of many face-related vision tasks such as face recognition.With the rapid development of deep learning technology,face detection and alignment have recently achieved state-of-the-art performance in both accuracy and speed.However,most previous methods focused on modifying the network structure to improve the performance of face detection or face alignment,while ignoring the high relevance of these two tasks.This paper focuses on the face detection and alignment using multi-task learning algorithms to improve the performance of face detection,and proposes a cascade regression algorithm combined with multi-task learning network to improve the effect of landmark positioning.Firstly,we design a face detection network based on feature pyramid structure,and add a facial landmark detection branch to it for multi-task learning.Since there is not any data set widely used in face detection and landmark detection at the same time,a facial landmark detection network is designed to automatically generate key point labels for face detection data set,and then the generated data set is applied to train the multi-task learning network.Due to the limited precision of the landmark label which is generated automatically,and in order to balance the loss of the two tasks,the designed multi-task learning network predicts the five key points of the face,and then apply the generated data set to train this multitasking network.The experimental results show that the designed face detection network and multi-task learning network have achieved good detection results,and the multi-task learning technology can effectively improve the effect of face detection.Subsequently,considering that the multi-task learning network only predicts the location of five landmarks of the face,which usually are not accurate enough,a cascade regression algorithm is used to further correct the results of key point location and to locate more key points at the same time.Since cascade regression is an iterative updating algorithm,the localization effect is easily affected by the initial shape quality,and it cannot locate landmark in complex scenes such as exaggerated expressions,extreme head postures,etc.,while the neural network is not susceptible to such disturbances.It is proposed to use the five key points of multi-task learning network detection to affine and transform 68 key points as the initial shape.Experiments show that the cascade regression algorithm initialized by multi-task learning network is more robust to complex face scenes.