A Method Of Face Detection Based On Facial Landmark

Face detection is one of the classic problems in compute vision literature.There are countless practical applications that require the application of face detection as the first step in their system.While the seminal Viola-Jones face detector can well solve the problem of face detection under constrained conditions,it fails when handling faces in natural images.In order to meet the demands of face detection in practical applications,researchers pay great attention to face detection under unconstrained conditions,and have made significant progress.Their works can be roughly classified into two general schemes: algorithm families based on cascade structure and algorithm families based on Deformable Parts-based Model,which yield significant improvement for face detection.More recently,convolutional neural network methods have made significant progress in image classification and object detection,and get state-of-the-art performance on these issues.Researchers also try to borrow them to handle face detection problems under unconstrained conditions.Again,they get state-of-the-art performance.However,these methods may be complex,difficult to train,or sometimes unstable.In this paper we present a new simple face detection algorithm,which obtains very competitive performance w.r.t the state-of-the-art methods on the FDDB benchmarks.The proposed method can be summarized as follows:First,the proposed method is a new simple face detection algorithm.This algorithm can be modelled as a single convolutional neural network and can be trained end-to-end.The proposed method is based on facial landmark.We produce face proposals by detecting the landmarks in the input images using a classification network and then combining the predefined boxes and K-means algorithm.Second,in order to improve the regression precision of the detected faces,we upsample the convolutional feature maps before landmarks detection by deconvolution to improve the resolution of these features.Multi-task loss function is also used to handle classification errors and regression errors properly.