High Resolution SAR Image Classification Based On Deep Feature Extraction And Conditional Random Field

With the rapid development of SAR image technology,it is possible to obtain massive high resolution SAR data.High resolution SAR data contains abundant information,with which more details in the image can be captured,enabling finely perception of the images.However,A variety of problems and challenges appear accompany with the continuous improvement of the SAR Image resolution,proposing higher requirement for SAR image understanding and interpretation.SAR image classification is the fundamental part as well as the key step of SAR image understanding and interpretation,aiming at assigning each pixel to a category according to extracted features.The accuracy of SAR image classification has a great impact on the subsequent stage of SAR image processing,thus,it is of great significance for the related research and becomes the hotspots of researchers.The main problems of high resolution SAR image classification includes:(1)SAR images are heavily contaminated by speckle noise,which harms the representation ability and robustness of the feature;(2)the semantic scene in the high resolution SAR images is complicated,leading the “semantic gap” between the image analysis and low level features.So,effective features are required to extract from the images before classification.Focusing on the problems of the high resolution SAR image classification,this paper studies high level feature extraction based on deep learning models and optimization strategy of conditional random field.On one hand,the high level features extracted from the deep learning models is able to solve the “semantic gap” of the high resolution SAR images,on the other hand,the high level features have great discriminant ability,they are robust and stable to reduce the influence of speckle noise as possible.The initial classification results with the deep learning model can be treated as a prior condition,based on which,the conditional random field model is able to combine the spatial information in the SAR images,further improving the classification performance.According to the above theories,two typical deep learning models: the stacked auto encoders and convolutional neural networks are analysised.The shortcomings of the two models includes: due to the restricted by the framework of the staked auto encoder,spatial information may be lost.In order to solve the problem,conditional random fields are introduced to take use of the spatial information of the SAR images.Besides,the framework of the convolutional neural network,making it can't take full use of the spatial information in the SAR images,aiming to solve this,recurrent neural network is introduced to encode the spatial information in the SAR images,improving the structure of the traditional convolutional neural network to achieve better image classification performance.Finally,based on the improved classification framework,a classification algorithm for high resolution SAR image classification is designed and completed.Experiments on three real high resolution SAR images show that the improved classification framework improves the classification performance and achieves stable high resolution SAR image classification.