Research On The Registration Method Of High-resolution SAR Image And Visible Light Image

During the last decades,remote sensing sensors have undergone a rapid development in data quality and characteristics.Although Multispectral Scanner(MS)device are informative in imaging,they cannot operate under complex natural conditions,such as nights,cloudy,or foggy weathers.SAR equipment is robust and can not be affected by these complex natural conditions.Due to the large difference in imaging principles of the two sensors,they reveal different characteristics of earth surface materials.Therefore,fusion of the two data can better explain the characteristics of imaging regions and provide better support for remote sensing applications such as change detection and disaster assessment.The premise of data fusion is image registration.Therefore,this dissertation focuses on registration with SAR images and optical images.Experimental results show that our proposed method is effective.The specific work of this dissertation is as follows:1)This dissertation studies conventional image registration methods and image registration methods based on deep learning.The development history and basic theoretical framework of image registration are studied.The similarity metrics commonly used in image registration,the geometric transformation function of images,the feature descriptors of images,and the evaluation criteria of image registration are summarized.2)Proposes a SAR image and optical image registration method based on structural similarity.In general,the registration of SAR image and optical image requires manual selection of control points(CPs),which does not consider the structural information of the multimodal images and can not implement a fully automatic image registration method.In this dissertation,first,we extend Harris detector's response function from linear item to quadratic form and build a new weight function by combining spatial and intensity information of pixels,which enable the location of corners more precisely.Next,we create a structural feature descriptor using both the amplitude and orientation of corners to provide more distinctive local image features.Finally,we set up the correspondence based on the generated point features,and map all pixels in the sensed image to the reference.Experimental results demonstrate that our improved detector can achieve much better detection performance compared with conventional Harris corner detector.In addition,registration with SAR and optical remote sensing images demonstrate the efficiency and accuracy of the proposed approach.3)Present a SAR image and optical image registration method based on deep learning architecture.Unlike conventional methods extract features and matching features separately,mapping function for registration can be learned directly from paired patches and their corresponding labels using deep learning framework.This end-to-end architecture allows us to take full advantage of limited data to optimize the whole model,which is not available with traditional methods.Experimental results show that the method is effective.