Application Of Texture Image Classification Based On Deep Learning In 3D Reconstruction

The multi-view three-dimensional reconstruction refers to the process of generating a three-dimensional geometric model from a two-dimensional image obtained from a general camera.In this process,the extraction and matching of feature points is one of the core techniques for obtaining high-precision 3D geometric models.Obtaining a sufficient number of accurate matching points is a prerequisite for 3D reconstruction.As the number of images increases and the resolution increases,feature point matching based on local features of the image can easily lead to mismatching,and even a correspondence between completely different images may be established.In the camera calibration or three-dimensional reconstruction problem represented by the Bundle Adjustment,the number of equations and unknowns increases exponentially as the number and resolution of images increase,and the number of images that can be processed at one time is limited.Usually,it is tens or hundreds,but for large-scale scenes,the original image can usually reach thousands or even tens of thousands.At the same time,due to lighting conditions,camera parameters,shooting angle,shooting distance and other factors,some images in the original images are inevitably unable to meet the basic processing conditions,and these images also need to be removed.Based on the above problems,this paper studies the effective classification,segmentation and grouping methods of original images.By analyzing the macro features larger than the local features,the original images are classified,grouped,segmented,etc.,which reduces the computational cost and improves the accuracy of feature matching,and improves the accuracy and robustness of the 3D reconstruction system.In terms of feature extraction,this thesis attempts to use the material texture dataset and the scene texture dataset to perform image texture classification training on the neural network for the first time,and compares it with the neural network trained on ImageNet.Based on the experimental results,the neural network trained on ImageNet was used for feature extraction.This paper proposes the following methods to solve the problem of texture image classification.First,reasonable cutting of images,and the cutting image is extracted and classified using a deep convolutional neural network.Then,each cutting graph feature vector is aggregated to make it a global vector representation of the original graph.Finally,the classification method of traditional clustering is used to solve the problem of texture image classification.The effectiveness of the strategy is verified in the specific application scenario of 3D reconstruction.The main research contents of this paper include:1)Introduce the main steps,related technologies and main problems of imagebased 3D reconstruction.Therefore,the rationality and necessity of texture classification preprocessing for 3D reconstruction of original input images are proposed.2)The method of image texture classification proposed in this paper and related conclusions are introduced in detail.It also provides new ideas for deep learning in the fields of image texture classification and image similarity classification.3)To facilitate management,browsing,and sharing of the resulting 3D model,a 3D presentation platform web project was developed using SSM(Spring?SpringMVC and Mybits)frameworks.