Research On Several Methods Of Face Detection Based On Cascade Structure

Face detection is the first stage of face analysis.It aims to confirm whether there is a face in the input image.If there is a face,the position and size of the face are output.In practical applications,imaging of human faces under uncontrollable conditions may be affected by various factors.Such as serious occlusion,uneven illumination,different expressions,low pixels and complexity of the background.All of these will bring higher challenges to the accuracy and speed of face detection so that face detection has been one of the main research topics in machine vision.In 2001,Viola and Jones have proposed a classic solution that makes it possible to detect frontal faces in real time with low computational complexity.Therefore,the framework has continued to be improved and widely used until recently.However,most of these methods rely on the accumulation of experience with high training costs.With the rapid development of the convolutional neural networks which enhanced expression of image features.In this paper,the discriminant projection Haar feature is proposed and the face classifier of Adaboost is trained using the improved cascade structure.Simultaneously,face detection algorithm based on cascaded convolutional neural network is proposed and compared with other method.The main work is as follows:(1)The current development status of face detection and challenges faced by face detection is analyzed.A detailed explanation of the assessment indicators such as the quality of the candidate box,the overlapping area,the recall rate and the accuracy rate is provided.(2)This paper start from the two important components of the face detection system design.Several classical frameworks of face feature extraction and classifier are researched in detail.Based on the Aggregation Channels Features.Fisher discriminant analysis is used to propose a projection Haar-like feature,which expands the channel features of the image and uses a soft cascaded Adaboost classifier.In the training process,the positive and negative samples were trained using the face structure information in the channel feature space to enhance the discrimination ability of the face features.The results were compared with other methods on the FDDB.(3)The classic framework of face detection based on convolutional neural network is studied in detail and a convolutional neural network structure is designed.The structure is divided into two stages.The first stage uses a low-pixel candidate window to input the shallow convolutional neural network to exclude a large number of background windows quickly;The candidate window passed from first stage will be adjusted to different scales by the image pyramid.In the second stage,the corresponding resolution images are respectively input into two branches of the network and the multi-scale convolution features are merged to output the final face classification and bounding box regression.During the training,the online training is conducted for difficult samples and the soft non-maximum suppression algorithm is used to perform multi-scale test on the datasets.The results were compared with other methods on two public datasets PASCAL FACE and FDDB.