Research And Application Of Face Detection By Concatenated Deep Convolutional Neural Networks

Face detection is a fundamental but important topic in machine vision and pattern recognition and has been extensively studied in recent decades.Face detection is one of the important steps in many face applications such as face recognition and facial expression analysis,and has important research significance.In a complex environment,face detection is affected the detection accuracy by a number of external factors,such as changes in face,posture,light and shielding.The paper proposes a cascaded deep convolutional neural network that the design of cascaded structure can locate more precise for face diagram,and deep convolutional neural network can extract more effective features to obtain better face detection performance.The main contents of this paper are divided into three parts: face detection algorithm in complex scene,face detection transplant based on Raspberry Pi platform,and face tracking of steering gear platform.(1)The algorithm research for face detection under complex scenes.Before training the network,the data needs to be processed.The proposed cascaded deep convolutional neural network model in this paper consists of three different convolutional neural networks groups.According to the input data size of each network,the images are processed into 12*12,24*24 and 48*48 respectively,and each input data contains positive and negative samples.Cascade deep convolutional neural network model training is mainly divided into three steps.First,a deep convolutional network is used to extract more effective features and a large number of facial candidate windows,and cross-entropy is used to discriminate faces or non-human face regions,and a large number of candidate windows of non-face regions are deleted.Then input candidate window that has not been deleted in the first step to the next layer network to further filter,and the remaining candidate window is input to the next layer network.Finally,output the final face window and to locate facial key position for candidate face.The cascaded structure not only can be effective to locate candidate windows of face region,but also be can improve the accuracy of face detection.Experimental results show that the proposed method has good performance of face detection,and can be widely applied to face recognition,facial key points and other related areas.(2)Based on Raspberry Pi Platform for Face Detection.Raspberry Pi is a small mobile development platform with good performance and rich hardware and software support,and is a popular development platform.Two kinds of face detection methods are transplanted to the Raspberry Pi platform.One is proposed cascaded deep convolutional network for face detection method in this paper,and the other is based on OpenCV of face detection method.Then to describe the detail face detection algorithm based on OpenCV.Finally,compared the two kinds of face detection effects on the Raspberry Pi Platform.Experimental results show that the proposed cascaded deep convolutional network method in this paper has good robustness under complex environment.(3)Face tracking based on steering gear platforms.The proposed face tracking method in this paper is to drive the steering platform to follow the face rotation so that the face can always be in the middle of the video.When the servo platform achieves the face tracking,computes the angles of rotate the servo platform needed according to the position of the human face in the video,and then the duty cycle of the PWM signal is calculated according to the angles of needed to rotate to achieve the human face tracking of the servo platform.The face tracking method based on the hardware platform can handle very well the problem that the camera can't capture the human face caused by people walking freely on video conferences.Therefore,the method can be applied to commercial fields such as conferences,video surveillance and et al.