Study On Deep Learning For Face Detection In Video

As the primary technology to extract information from face,face detection is the premise and foundation of face recognition,facial expression analysis and face tracking.At the same time,face monitoring has a significant advantage that it is not easy to be found by the monitored object.Thus,with the expansion of video surveillance coverage,video face detection has been more and more used in crime analysis,intelligent security,and artificial intelligence.However,many factors need to be considered in video face detection,such as the complex environment,the partially obscured face,the face rotation angle,and so on.In addition,the detection speed is also need to be taken into account because of the real-time requirement.Point to these influence factors,a video face detection method based on the deep learning and the video's inter-frame continuity is proposed,which is trying to achieve stronger robustness against complicated backgrounds,illumination,rotation angles with the lower missing detection rate and the lower false detection rate,and provide theoretical support for intelligent monitoring and wisdom security.The concrete research content is as follows:First of all,according to the theory of deep learning and face detection based on neural network,a cascaded learning network based on multi-layer Probability state-Restricted Boltzmann Machine(PRBM)is proposed to realize face detection in a single frame video image.It first uses the probability state of the neurons in P-RBM as their activation degree,which better models the activity state's continuous distribution of the neurons in human brain.This design not only retains the weak active information,but also decreases the effect caused by the former layer's error.Secondly,this method simulates the hierarchical learning mode in human brain by cascading multiple P-RBMs.This cascaded network can realize the multi-layer nonlinear mapping and obtain the semantic feature of the input date.What's more,it can learn the relationship hiding within the data to make the learned features be more promotional and expressive.Simultaneously,the number of the hidden layer's neuron decreases layer-by-layer to control the network's scale and enhance the robustness.Finally,it uses the layered training and whole optimization to balance the robustness and accuracy.This face detection in the single frame image method does not use the inter-frame continuity information,which is the unique advantage of the video.Thus,based on the above research,the continuity between video frames is to be further studied,and a video face detection method with multi-inter-frame information fusion is proposed.Firstly,the aspect ratio of the face skin color area is used to remove some mistakenly detection areas.The threshold of the aspect ratio is set by an adaptive update method in order to obtain the most appropriate boundary condition for the detection video.Subsequently,the change rule of the face location between video frames is used to estimate the detection result in the current frame.Then,the estimate result is compared with the real detection result,and a contrast rule is used to modify the detection result of the deep learning network according to the contrast difference,deleting the false detection area,filling the missing detection area,which improves the detection accuracy.The experimental results show that,no matter static single face detection or multiple faces detection under complicated conditions,besides the faster detection speed and stronger robustness against face rotation,the cascaded P-RBM learning network possesses the lower false detection rate and the lower missing detection rate.Moreover,combining it with the multi-inter-frame information fusion method to detect face in video not only keeps the faster detection speed and the lower false detection rate,but also reduces the missing detection rate significantly.In addition,it improves the detection performance of the partially obscured face.