3D Object Reconstruction And Recognition Based On Deep Neural Network

3D reconstruction based on image sequence is an important research filed in computer vision.Compared with two-dimensional images,three-dimensional data can make people understand the spatial geometry information of the target more intuitively.3d object reconstruction and recognition is the key technology of 3d scene understanding,and the basis of computer understanding and interaction with the world.It is widely used in the fields of virtual reality,augmented reality,intelligent robots and video games.In this thesis,the research is carried out from the two directions of 3d target reconstruction and 3d target recognition.The specific work can be summarized as follows:At present,Structure from Motion(SFM)is used to reconstruct the 3d structure of the scene object in a series of 2d images by hand design feature extraction.However,in the face of targets with weak textures or light changes,target reconstruction is prone to problems such as holes and information loss.Aiming at this problem,this thesis proposes a new algorithm for predicting depth maps of fully convolutional networks based on multi-scale feature fusion(FCNMFF).The new algorithm introduces the idea of feature pyramid network,and uses multi-scale learning strategy to estimate the depth image with richer spatial structure details.Finally,the depth map estimated by the new algorithm is used for fusion reconstruction.Experimental results show that this method can effectively solve the problem of 3D reconstruction of weak texture targets.The ultimate goal of 3d object reconstruction is to realize object recognition and classification.In terms of 3d target recognition,based on multi-view representation of three-dimensional data,this thesis proposes a multi-view feature fusion algorithm based on recurrent neural network.The model uses the recurrent neural network to construct the association information between multi-view images,and finally uses the Weighting method to perform multi-feature fusion to generate a three-dimensional target shape descriptor,thereby fully mining the association information between the multi-view images.Experimental results show that the new algorithm has higher recognition accuracy than some advanced 3D target recognition algorithms.Finally,in the 3D target recognition experiment,this thesis proposes a 3D target recognition scheme based on multi-view depth estimation and hybrid view fusion,which further improves the 3D target recognition rate.