Research On SAR Image Registration Method Based On Deep Auto-Encoder

SAR images have the advantages of all-weather work,strong penetrating power,etc.They are widely used in meteorological analysis,geological survey and other fields.SAR image registration is one of the essential steps in many SAR image processing applications.Therefore,the research of SAR image registration technology has very important research value in today's society.Because of its special imaging mechanism and imaging conditions,SAR images have speckle noise.However,the registration methods of traditional optical images cannot overcome the influence of noise and make them unsuitable for SAR images.Therefore,registration methods based on SAR images have become one of the hotpots in the area of image processing,and has been rapidly developed.:Supported by the National special support program for high-level talents of China(The interpretation and object detection of SAR image),and the National Natural Science Foundation of China(Pol SAR image classification based on generative adversarial network,No.61771379).This paper propose three SAR image registration methods using deep auto-encoder network,which base on feature-based image registration method and combine with deep learning.Main tasks are as follows:First,we explain the problems of SIFT features in SAR images through experiments.For example,SIFT features no longer have rotational invariance and do not satisfy nearest distance matching.For these questions,we extract image features unsupervised based on auto-encoder network.The input of the network is an image block centered on the feature point,and the output is used as a feature description of the image block.Then the features are measured to get matching pairs,and complete the registration.This method avoids the complex operatio n of artificial design features,and for different data,it can adaptively extract features for matching.Compared with SIFT,about 1.5 times the number of correct matching pairs can be obtained,and RMSE is smaller,thereby improving the accuracy of registration.Second,since the auto-encoder network encodes and reconstructs image data as an unsupervised process,the updating of network parameters is not related to the matching results of the features.Therefore,in practical applications,it is easily that the features extracted by the similarly-shaped image blocks are also similarly.To solve the problem,we trained the auto-encoder network through supervise learning.At the same time,in order to better learn data distribution,and effectively use historical data,we introduce transfer learning.We used historical data for pre-training,and target data for fine-tuning.After training,the matching results of the feature points in the test image can be directly obtained without distance measure.This method can obtain a greater number of correct matching pairs and reduce RMSE,further improving the accuracy of registration.Third,supervise learning includes two processes: pre-training and fine-tuning.Compared with auto-encoded networks which is unsupervised,the accuracy of registration has been improved.But the training process is cumbersome and the training time is too long.In order to improve the e fficiency of training,this paper uses a training method which combines pre-training and fine-tuning,and it based on the training process of supervised learning.That mean,the cross-entropy and reconstruction error are combined to construct the loss function,and different weights are used to determine the role of each part at different stages of training.Compared with the training process of supervised learning,this method simplifies the training process.And at the same time,it ensures the number of correct matching pairs,and then ensures the accuracy of registration.