Research On Key Technologies Of 3D Moving Object Reconstruction Based On Deep Convolution Neural Network

Compared with other human perceptions,vision is the most reliable source of information,and the information received by human through vision accounts for more than three quarters of the total information acquired by human beings,so visual information is very crucial.Image deblurring and three-dimensional reconstruction have a wide range of applications.In recent years,it has become one of the hot issues in the field of image and three-dimensional research at home and abroad.To solve the problem of image deblurring,a prior-driven image deblurring depth network is proposed to combine the advantages of optimization-based and discriminant learning-based image deblurring methods.Firstly,an image deblurring method based on denoising is established,the iteration process can be carried out efficiently.Then,the iteration process is expanded into a feedforward neural network.The network layer of the neural network simulates the flow of the image deblurring algorithm based on denoising.In addition,an efficient CNN denoiser using multi-scale redundancy is built and inserted into the deep network.Through end-to-end training,the CNN denoiser and other network parameters can be optimized jointly.The experimental results show that the proposed method is very competitive in the performance of image deblurring.To solve the problem of three-dimensional reconstruction,inspired by the success of long short-term memory network and the latest progress of single-view three-dimensional reconstruction using convolutional neural network,this thesis establishes a structure based on shape prior,which is called three-dimensional reconstruction recurrent neural network.Shape prior-based methods can use fewer images and have fewer assumptions about object reflection functions.The architecture is based on standard LSTM and GRU and consists of three parts: two-dimensional convolution neural network,three-dimensional convolution LSTM unit and three-dimensional deconvolution neural network.The network receives one or more images of an object instance from different viewpoints and outputs the reconstruction of the object in the form of three-dimensional occupied grids.In training and testing,the network does not need any object class labels or image annotations,and it only needs little supervision.A key attribute of the network is that it can selectively update hidden representations by controlling input gates and forgetting gates.In training,the mechanism allows the network to learn the appropriate three-dimensional representation of objects adaptively and consistently,so conflicting information from different viewpoints can be utilized.When traditional SFM or SLAM methods fail to reconstruct images sequences,this method can reconstruct images with insufficient texture or wide baseline viewpoint.