Research On The Training Methods Of Feedforward Neural Network Based On Nonlinear Filter Optimization

3D reconstruction has become a research hotspot in many disciplines such as computer vision,image processing and virtual reality technology in recent decades of development,and it has been widely used in the game industry,construction industry,historical event and other fields.3D reconstruction is to obtain the 3D model of the object or scene.3D reconstruction based on multiple views is one of the main means to obtain the 3D model and point cloud reconstruction is an important process of it.The core algorithm is based on multiple images of an object or scene from different angles as input,and for e ach image for feature extraction and matching to obtain three-dimensional point cloud data,and finally obtains the final 3D model by surface reconstruction.But this kind of algorithm is too complex,and the image background contains a large number of redundant information,it requires a large amount of calculation in the 3D point cloud data,and moreover it is not real-time,and t he reconstruction of different types of images have different effects.So to improve the reconstruction's effect and its real-time performance has become the key problem to be solved in multi view in 3D reconstruction.This paper is based on 3D reconstruction algorithm which is based on patch proposed by Furukawa and then launches the research on the ancient architecture 3D Reconstruction Technology(Patch based multi-view Stereo,PMVS)relied on the laboratory project.The main research work is as the follows:This paper first in troduces the existing 3D reconstruction methods,aiming at t he structural characteristics of ancient architecture,reconstruction effects of feature extraction getting from Harris and DOG algorithm in PMVS is not satisfactory.In order to solve this problem,this paper proposes a PMVS algorithm based on saliency region detection SIFT(FT-SIFT).The ex perimental results show that the feature extraction algorithm based on FT-SIFT is better than that of the original PMVS algorithm,it can also remove most of the background redundancy feature points,greatly reduce the time to generate the feature descriptor.The algorithm not only guarantees the reconstruction effects of the point cloud data,but also improves the reconstruction efficiency.Secondly,in order to make the final model more close to reality,the number of initial input images of 3D reconstruction is relatively high,the amount of calculation is increased,and the time is longer.For multi view 3D reconstruction is not real-time,so the paper first analyze the process of 3D reconstruction of point cloud,and then introduces the concept of GPU to solve the problems of time-consuming such as feature extraction and matching.On the premise of ensuring the accuracy and efficiency of the reconstruction,the improved FT-SITF algorithm is optimized by GPU in the CUDA framework.The experimental results show that compared with the original PMVS algorithm,the improved PMVS algorithm of FT-SIFT improves the precision and ef ficiency of poi nt cloud reconstruction,and a chieves the desired effect by GPU parallelization.