Research And Implementation Of 3D Objects Reconstruction Based On Recurrent Neural Networks

This paper makes full use of deep learning related technology and theoretical knowledge,studies the working principles and advantages of densely contected,convolutional and recurrent neural network,and sums up the shortcomings and defects of 3D-R2N2 recurrent neural network.On these bases,this paper presents and implements a series of improvements to 3D-R2N2 network.And the improved model outperforms the original version in all aspects,such as speed,space occupancy rate as well as reconstruction accuracy.The specific work done in this article is as follow:This paper puts forward and realizes the improvement of the 3D-R2N2’s Encoder module.The improved model has replaced normal convolutional layers with densely connected version.Compared with the old model,the new one can be trained faster and more steable,and the loss function is able to converge faster,at the same time,the new model has higher reconstruction precision.This paper puts forward and realizes the improvement of 3D-R2N2’s loss function.In the improved model,a new loss function based on chamner distance(DT)has substituted the original cross entropy loss function.By using new loss function,the training speed of proposed model is faster,it also helps proposed model to obtain smaller loss value and have higher reconstruction precision.This paper improves 3D-R2N2’s Decoder module.In the Decoder module,this paper adds a new branch network to analyze and extract color related information,and this branch network can also color the reconstructed 3D models according to the color information provided by images.The 3D model structure is more clear after the coloring processing,so the user can observe and evaluate the reconstructed model more clearly.In this paper’s experiment,the improved model has been tested on ShapeNet and PASCAL 3D databases.At the same time,this paper also tests the original 3D-R2N2 and other two deep-learning-based 3D reconstruction methods.During the whole experiment,this paper uses information visualization tool to analyze and summarize the data generated in the experiment.And proposed model is better than others.