Research On SAR Image Processing Based On Deep Learning

Synthetic Aperture Radar(SAR)has always played an important role in the military and civilian sectors.With its advantages on huge information acquisition and all-weather work,it has a wide range of applications in resource detection,target tracking,and danger warning.With the continuous development of remote sensing technology,synthetic aperture radar technology is more and more perfect,and has great application value.In many research fields,research on SAR images(target recognition classification,super-resolution reconstruction,target segmentation,etc.)has always been the focus of research.For these tasks,feature extraction is a key factor in determining the performance of the algorithm.How to select effective features becomes the core problem of the task.In recent years,various image processing algorithms have developed rapidly in the field of optical imaging,and many new model concepts have emerged.Such as the spatial pyramid on the feature expression,the sparse representation concept,the character bag model on the character feature processing,and so on.These models and algorithms excel in the task of optical imaging,and people have begun to introduce these excellent algorithms and concepts into SAR image processing tasks.In 2006,Hinton et al,In the idea of deep learning,it is advocated to automatically learn the characteristics of images by constructing a multi-layered deep network model.As the depth of the network deepens,the learning characteristics also change from shallow to deep.Experiments show that the features extracted by the deep learning network perform better in various image processing tasks.At this point,research on deep learning has become in full swing.This paper introduces the concept of deep learning into the SAR image processing task,mainly to study its two tasks in the processing task:classification and regression.Thereby the two tasks of target recognition and super-resolution reconstruction of SAR images are completed.However,due to the multiple characteristics of SAR images,the study of deep learning SAR image processing has been hindered:(1)Unlike optical images,SAR images by microwave sensors and then records the echo information of the objects in binary complex form.These information need to be formed into images that can be studied and analyzed by later imaging algorithms.The imaging algorithm affects to some extent.The image quality.In addition,SAR images contain a lot of unavoidable speckle noise,which makes the target easy to be submerged in noise,and it is difficult to extract valuable information.Therefore,SAR images are not easy to understand and compile compared to optical images.(2)The excellent performance of deep learning is based on a large number of sample training.In the training task of optical images,the training set often contains tens of thousands or even hundreds of thousands of samples.However,due to various factors,SAR images have not yet had a large public training set as optical images.And in certain engineering tasks,due to time,economic benefits and other factors,the adoption of large SAR image training sets has become more difficult to achieve.How to implement sample extension on existing SAR images has also become a big problem.In order to introduce deep learning into SAR image processing tasks,this paper focuses on the classification problem—SAR target recognition,regression problem—super-resolution reconstruction.Combined with the above difficulties,the following work is carried out:(1)For the problem of insufficient SAR image samples when using the deep learning method,a SAR image expansion technique is proposed.According to the imaging principle and imaging quality of SAR images,relevant pre-processing operations are carried out to achieve enhancement in multi-dimension.The processed SAR image is combined with the original image to form high dimensional data as a network training sample.It is beneficial to improve the problem of small sample size of SAR image,so that the data can support network training,thereby automatically learning the features more suitable for classification,so that the recognition accuracy reaches the expected level.(2)Due to the low resolution of SAR images,many targets are difficult to distinguish in multi-target images,and SAR super-resolution reconstruction is also crucial.Combined with the current multi-resolution reconstruction network,the network structure,loss function and training strategy are improved for specific engineering requirements,so that the trainable network can obtain better high-resolution SAR images.On the basis of solving the above theoretical and technical problems,the thesis studies a variety of deep learning networks and training strategies.The experimental results show that deep learning has great potential in the field of SAR image processing.