Research On Super-resolution Reconstruction Of Sequential Images And Its 3D Reconstruction Method

With the prosperous development of computer vision technology,only obtaining the two-dimensional information of the target in the scene through the computer gradually cannot satisfy people.The process of using two-dimensional sequence images to reconstruct the three-dimensional information of the target subject has become an increasingly urgent need for people.Nowadays,the 3D reconstruction technology based on sequence images has played an irreplaceable role in multiple areas,but this technique can not give consideration to the reconstruction accuracy and real-time performance.This thesis mainly focuses on the accuracy and realtime of 3D reconstruction of sequence images.It starts with the research on these two issues,which require high resolution of input sequence images for 3D reconstruction,and the inability of traditional 3D reconstruction algorithms to take into account accuracy and realtime.Firstly,in view of the two characteristics that the 3D reconstruction algorithm requires high resolution of the input sequence image and the existence of a large amount of similar texture information between the adjacent images of the sequence image,this thesis proposes an image super-resolution reconstruction network CANTT(Image Super-Resolution by Channel attention and Neural Texture Transfer)based on channel attention mechanism and neural texture migration.Through experiments,it is proved that compared with several classic image super-resolution reconstruction networks,this network can obtain more rich texture details and real reconstruction images,which lays a foundation for the 3D reconstruction algorithm based on sequence images to obtain high-precision 3D models.Secondly,in view of the existing mainstream 3D reconstruction solutions that often fail to balance real-time and accuracy,this thesis proposes an image feature matching model based on the SIFT-GMS algorithm,and designs a solution based on this model to realize the 3D reconstruction of sequence images.The image matching model proposed in this thesis achieves good real-time on the basis of ensuring the reconstruction accuracy.The time consumption of this model is only 34.6% of the SIFT algorithm and 74.1% of the SURF algorithm,while showing a matching effect that is not inferior to the SIFT algorithm.It is an algorithm that can balance real-time and accuracy.At last,this thesis uses the adjacent next image of the sequence image as the reference image of the previous one,and uses the CANTT network proposed above to improve the image quality,and then uses the high-resolution sequence images obtained by super-resolution reconstruction to perform 3D reconstruction to significantly improve the accuracy of the final3 D model.This study provides a practical idea for how to improve the accuracy of image-based 3D reconstruction Algorithm.