Design And Implementation Of 3D View Fusion Algorithm Based On Deep Neural Network

Multi-view stereo vision technology is a very reliable and practical technology to generate 3D point cloud image model by processing the related information of multiple related 2D images through neural network.With the rapid growth of computer computing power,which has led to the rapid development of related technologies in the field of deep learning,the promotion of 3D view fusion technology is also visible to the naked eye.However,the high cost of technology,manpower and time brought by traditional 3D reconstruction technology,as well as the low efficiency,low quality and even dangerous view acquisition methods of required objects lead to unsatisfactory generation effect.Therefore,after a large number of in-depth studies on various 3D view fusion technologies and methods at home and abroad,this paper integrates related technologies and improves them,and designs a set of 3D view fusion algorithm based on deep neural network for view fusion reconstruction of objects.Specific research contents and innovation points are summarized as follows:1.Conduct in-depth research on the field of deep learning.Analyze the construction principle of convolutional neural networks,the technical principle of generative adversarial networks,the relevant technology of traditional 3D reconstruction,and the principle and application of the latest 3D view fusion technology based on deep learning.Compare the advantages and disadvantages of these approaches and summarize the main research direction for this paper.2.This paper presents a design for an improved HTHGAN network-based super resolution enhancement technique for low resolution images.In real-life scenarios,image pixels collected from objects are often affected and interfered by various factors,which can impact the extraction and processing of image features by subsequent 3D view fusion models.In this paper,an improved HTHGAN view super resolution model is proposed.A dense connection residual module is embedded in the "low to high" network layer generator to improve the density of feature extraction,so as to improve the resolution and enhance the image sequence one by one.The results of follow-up experiments show that the super resolution image output by the method designed in this paper is good and meets the standard of later work.3.This paper proposes a multi-view stereo vision network M-MVSNet based on multi-scale feature extraction module.After in-depth analysis and research on various multi-view stereo vision networks,it is found that the accuracy of 3D view fusion is directly affected by the finetuning of extracted image features.Based on the classic multi-view stereo vision network MVSNet,a multi-scale feature extraction network M-MVSNet is proposed.A new lightweight multi-scale feature extraction module is introduced to extract dense feature information and avoid the loss of key information in the image.The proposed method can connect the multi-scale feature information of the image while maintaining the original resolution of the image,so that the final fused 3D view is more accurate and complete.After a series of experimental comparisons,the 3D view integrity and accuracy of the proposed method for 3D view fusion and the image quality are higher than other methods,which proves the advantages of the proposed method.4.Based on the above method,a simple and practical 3D view fusion system is designed and realized.Firstly,the architecture of the system is briefly outlined,and then the function module is designed according to the content of this paper.After the design of interface,system interface and system function module,this paper will implement the visualization of this system.The system according to the actual demand analysis,after a variety of system testing methods comprehensive evaluation system.The experimental results finally show that the system is efficient and practical,and the system has a certain scalability,the realization of three-dimensional view fusion effect,so that the algorithm proposed in this paper has been applied in practice.