Research On Regression Model Based On Deep Neural Networks And Its Applications

Regression model is the most popular mathematical model used for prediction.The model can be designed in different research fields,such as object detection,face recognition,human action recognition and so on.The traditional regression model can fit non-linear functions by learning a small number of labeled samples,but there are huge challenges for it to learn complex mappings and extract robust features when these problems are different and there are massive amount of data.Since deep neural network as an advanced technique can extract features from data automatically and learn complex mappings between input data and output labels,we use the regression model based on deep neural networks to reduce the influence of the above problems for predicting and improving forecast accuracy.This dissertation studies the research on the regression model based on deep neural networks from two aspects,including the constructed regression model based on deep neural networks and regression model based on deep neural networks feature.Because head pose estimation is classical regression problem and human action classification is generalized regression problem in computer vision,different regression models are designed based on deep neural networks for these tasks.The regression models not only improve the accuracy,but also provide insights for other tasks.The main studies and contributions are summarized as follows:1)In order to overcome the limitations that the traditional methods have strong dependence on initialization and accurate facial landmark location,a cascaded deep neural network method is proposed for head pose estimation.The proposed method presents a cascaded deep neural network framework and a multi-level regression algorithm to estimate head pose gradually.Firstly,a regression model based on a cascaded deep neural network is designed to estimate human head pose and the deviations of head pose.Then,a multi-layer regression algorithm is introduced to obtain the final estimation by a coarseto-fine way.The regression model based on cascaded deep neural network makes the estimated pose approximate the ground truth step by step,and the multi-level regression algorithm computes the estimators from different layers to obtain the final estimation.The experimental results show that our method can estimate head pose from the facial images without initialization and accurate facial landmark location.2)For the sake of modeling the correlation among the degrees of freedom of head pose,a scale invariant constrained network is proposed for head pose estimation.The proposed method further reduces the mean absolute error according to different pose change law.This method firstly extracts global and local features of the facial images,and then it estimates head pose and the deviations of head pose through the global and multiple local network layers.Finally,the initial estimation successively adjusted according to the estimated deviations.Furthermore,by analyzing the characteristics of the facial images under different individual with different head pose and the same individual with different head pose,the empirical evidence is obtained for our model.The experimental results show the effectiveness of the method.Additionally,it is found that the proposed method is preferable to estimate human head pose,but its testing time is four times for the cascaded deep neural network method.Therefore,this method is applicable for general requirements for speed but higher requirements for accuracy.3)In order to overcome the disadvantages that the existing methods have pool results under the condition of inadequate,incomplete training data,a pose sensitive multivariate label distribution learning method is proposed for head pose estimation.The proposed method adopts a deep neural network to extract robust pose features and builds a multivariate label distribution learning algorithm for head pose estimation.Firstly,the ResNet is used to extract robust pose features of the facial images.Then,multivariate label distribution learning algorithm is employed to estimate distribution of the facial images,and compute the final estimation according to this distribution.In multivariate label distribution learning algorithm,a multivariate label distribution is generated for each head pose to describe the importance of each pose for the facial image by using multivariate Gaussian distribution.This can not only alleviate the problem of inaccurate pose labels,but also boost the training examples associated to each pose without actually increasing the total amount of training examples.The experiment results show that the proposed method can estimate the head pose effectively under the circumstance of insufficient and uneven training samples.4)In order to make fully use of the prior knowledge and make the feature semantics more clear,a new human action classification method based on two kinds of recurrent neural networks and a two-stage regression model is proposed.In this method,echo state networks are utilized to extract local features which can provide a useful geometric characterization for human actions,and the prior knowledge of adjacent area for the twostage regression model is introduced to classify human actions.Firstly,the echo state networks are adopted to obtain the conceptors of the trajectories of each dimension of skeletal joint point coordinate.Secondly,the local features of human action sequences based on these conceptors are computed.Then,the first stage network is used to predict fusion features of each human action.Finally,by using the second stage network,human action is classified accurately based on the fusion features.The experimental results on different datasets show that our algorithm can classify human action effectively in complex scene.