Research On Remote Sensing Image Calibration Algorithm Based On Descriptor And U-Net

After decades of development,artificial intelligence technology has developed rapidly,and image-processing technology has become more and more mature.As an important branch of image processing,image registration has attracted extensive attention in academia.At the same time,image registration has also been widely used in many fields,such as machine vision,medical science,geographic remote sensing,and other fields.However,with the high complexity of 3D image optimization,it is difficult to define the difference between the reference image and the image to be registered.If there is no pixel-aligned data,it is difficult to train the Generated Adversarial Networks(GAN)to generate a composite image,where all textures are accurately mapped to the reference image.At present,many works have proposed remote sensing image automatic registration technology,but the performance of processing different images has not been comprehensively studied to analyze these technologies.Because of the above problems,the main research points of this thesis are as follows:1.Remote sensing images taken under different conditions may show nonlinear changes.The registration of these images is an important process.Descriptors have unique characteristics,mainly reflected in their robustness to changes in observation conditions and sensor errors.The detector and descriptor are separated in this thesis.Their performance depends on the interest point detector used.2.An optimization of the SURF algorithm for color images,called O-SURF,is proposed.It is a combination of the MSER detector and the SURF descriptor.This method uses the MSER detector to extract the set of interest points and uses the SURF descriptor to describe it.The computation cost of the O-SURF descriptor can be reduced by implementing parallel computation.The experimental results show that the proposed improvement is effective compared with the basic SURF version.3.Based on the U-Net registration network of geographical remote sensing images,this thesis improves a U-Net image registration network model that combines semi-supervised antagonism and attention mechanism.This method is composed of the decoder and encoder,which are mainly based on sampling and convolution.The feature information of the encoder is integrated by the addition operation,and then the spatial attention mechanism and the channel attention mechanism are combined to introduce between the convolution layer and the pooling layer.The fused features can suppress the irrelevant areas in the geographical remote sensing image,highlighting the significant features of geographical hyper-remote sensing;Finally,the final image registration is achieved by combining spatial transformation and anti-line similarity optimization.4.The U-Net registration method of semi-supervised antagonism and attention mechanism is evaluated in the geographic remote sensing data set.The simulation results show that this method has a great improvement in registration accuracy and speed compared with other classic registration methods.This thesis compares the five descriptors of SIFT,GLOH,SURF,PCA-SIFT and SURF-HISTEQ.The proposed image registration technology aims to solve the classical problems of accuracy,robustness,speed and automation in remote sensing images.The experimental results show that the O-SURF improvement proposed in this thesis is effective compared with the basic SURF version.Based on the multimodal geographical hyper-remote sensing image U-Net registration network,an improved method of combining semi-supervised antagonism and attention mechanism of the U-Net registration network is proposed.Our method is evaluated in geographical data sets,and the experiment shows that compared with the registration method of deep learning,this method has great improvement in registration accuracy and efficiency.