SAR Image Denoising And Change Detection Based On Deep Network And Extreme Learning Machine

Synthetic Aperture Radar(SAR)with characteristics of all-time and all-weather,finds wide application in the people's livelihood and military.And as an important data source,SAR image has aroused great concern.However,SAR image is corrupted by speckle noise with the coherence of scattering phenomena.Though speckle noise carries some of the noisy area information,it seriously damages the space information of the image,which subsequently affects the tasks of image segmentation,classification,recognition and change detection.Thus,it makes great sense to develop appropriate filtering technology to suppress speckle noise.Also,SAR change detection is in wide use recently.In this paper,SAR image denoising and change detection methods based on deep network and extreme learning machine are studied.The main works are as follows:(1)A SAR image denoising method based on residual learning and deformable convolution network is proposed.The residual learning mechanism is used to learn the mapping from the noisy image to the residual image between the noisy image and the clean image for the ease of network training.And the method replaces parts of the original convolution layers with the deformable convolution layers.The deformable convolution layer learns the shape of convolution layer with the reduction of the loss function,which consequently makes the model fit the actual problem better.The proposed method combines the residual learning and the deformable convolution network to deal with the SAR denoising problem.Experimental results show that this method gets better denoising results in both simulated SAR images and real SAR images.(2)A SAR image change detection method based on saliency mechanism and deep convolutional network is proposed.The saliency areas in an image usually have greater difference over other areas,which can highlight the information of changed areas.Thus,we adopt the saliency mechanism to further increase the difference between changed areas and unchanged areas.And in this method,a deep convolutional network,which extracts more advanced and more abstract featrues and has a strong classification ability,is established.The network then divides the processed log-ratio difference image into changed areas and unchanged areas,and the final change detection result obtains.Also,a novel pre-classification method and selection strategy of the training sample are presented to improve the classification accuracy of the deep convolutional network.Compared with some existing change detection methods,the proposed method obtains great results.(3)A SAR image change detection method based on multi-level feature learning and extreme learning machine is proposed.The sparse autoencoder has strong unsupervised lerning ability which extracts deep characteristics of the internal structures of the data.And the multi-layer extreme learning machine not only has strong ability of feature extraction and classification,but can learn fast.Hence,this method uses several sparse autoencoders to learn multi-level features from the input images and combines with the multi-layer extreme learning machine for the final feature extraction and feature classifiction.And obtains the result of change dectection in the end.Experiment results show that the proposed method gets great change detection results.