Continuous Disparity Dense Matching Of Satellite Images Based On Neural Architecture Search

When the image pairs of the target region cannot be acquired by stereo mapping satellites in time,it is one of the important technical means to acquire the large-scale high-resolution DSM quickly by 3D reconstruction using incidental satellite image stereopairs.This is of great significance to emergency surveying and mapping.Due to the limited feature extraction ability of traditional dense matching methods,it is difficult to deal with the radiation and geometry differences of incidental satellite images,and it is also difficult to obtain well matching results in regions weakly textured,no textured and highly reflected.Therefore,it is a problem worth to be studied that how to reconstruct 3D structure from incidental satellite images automatically and accurately.Deep learning technology can extract deeper features from images and has strong cost volume regularization ability in dense matching tasks,therefore,deep learning stereo matching model has attracted extensive attention in the field of computer vision.Because there is biggish difference in radiation and angle of view between pushbroom satellite images and frame images,these stereo matching models may not play their full potential in the stereo matching tasks using high-resolution incidental satellite images.Designing a new deep learning model for dense matching tasks using high-resolution incidental satellite images requires a lot of expert knowledge and has some limitations.In order to improve the effect of incidental satellite images dense matching and three-dimensional reconstruction,this paper has attempted to use neural architecture search technology to automatically search the best network architecture on the incidental satellite images.The main research work is as follows:(1)Satellite learning effective architecture stereo model(Sat-LEAStereo)has been constructed.In view of the fact that designing a new deep learning model for high resolution satellite images matching tasks requires a lot of expert knowledge and has some limitations,this paper has introduced neural architecture search technology into the high-resolution satellite images matching tasks.The construction method of cost volume has been modified to adapt to the conditions that positive and negative disparity exist simultaneously.Satellite learning effective architecture stereo model Sat-LEAStereo has been constructed,and the optimal parameters of candidate operations and network layers have been determined by comparative experiments.(2)Satellite continuous disparity learning effective architecture stereo model(SatCDLEAStereo)has been constructed.Aiming at the problem that the disparity regression module of Sat-LEAStereo is easy to cause the adhesion of foreground and background at the parallax fracture,the continuous disparity module has been introduced to enable the network to directly output sub-pixel disparity without disparity regression.Aiming at the problem that the smooth L1 loss function cannot reflect the pixel neighborhood information,the Wasserstein loss function has been introduced to guide the model to learn the real disparity distribution.And the satellite continuous disparity learning effective architecture satellite stereo model Sat-CDLEAStereo has been formed.(3)The three-dimensional reconstruction method using incidental satellite images based on neural architecture search has been studied.Taking the urban semantic 3D dataset as an example,the experiment of 3D reconstruction using incidental satellite images based on SatCDLEAStereo model has been carried out.Sat-CDLEAStereo proposed in this paper has been used to complete the intensive matching task in the process of 3D reconstruction,and compared with other deep learning methods and traditional methods to verify the feasibility and superiority of Sat-CDLEAStereo.It is shown that,(1)The network architecture with better performance can be obtained by using neural architecture search technology.(2)Compared with other deep learning models designed manually and traditional methods,the stereo matching results with higher accuracy can be acquired by SatCDLEAStereo with continuous disparity module.And the continuous disparity module can effectively reduce the matching error at the disparity fracture.(3)Compared with other deep learning models designed manually and traditional methods,the application of Sat-CDLEAStereo in the three-dimensional reconstruction using incidental satellite images can acquire DSM with lower mean absolute error and root mean square error.The three-dimensional reconstruction method using incidental satellite images based on continuous disparity learning effective architecture has a good effect in the three-dimensional reconstruction using incidental satellite images,and is expected to better serve the work of emergency surveying and mapping.The research work of this paper may have a certain reference value for the application of neural architecture search technology in threedimensional reconstruction using satellite images.