High Resolution SAR Image Classification Based On Curvelet DLN

SAR(Synthetic Aperture Radar)image classification is the basis of SAR image processing and interpretation.With the increase of SAR image resolution,the interpretation of SAR image poses new problems and challenges,such as sharp increase in data volume,richer and more detailed features,difficult to obtain artificially labeled samples,etc.Most current methods for SAR image classification are no longer suitable for high-resolution SAR images.Therefore,how to effectively classify high-resolution SAR images has become a research focus in recent years.Based on the idea of semi-supervised deep learning,this paper proposes a high-resolution SAR image classification method based on deep ladder network.The model can make full use of a large number of unlabeled data and a small amount of labeled data to extract discriminative features.The deep ladder network is applied to different high resolution SAR images to achieve higher classification accuracy.Our contribution in this paper are as follows:1.A high-resolution SAR image classification method based on deep convolutional ladder network is proposed.This method is a kind of semi-supervised classification method,which can make full use of a large number of unlabeled data and a small number of labeled data for feature extraction.And this method optimizes both the supervised loss function and the unsupervised loss function,guiding the network to learn more discriminative features from unlabeled data and improve the classification accuracy.The deep convolutional ladder network is used to classify the five groups of different SAR images.The analysis of the experimental results shows that this method obtains higher classification accuracy than the methods such as the stack autoencoder and residual network.2.A high-resolution SAR image classification method based on curvelet deep ladder network is proposed.Firstly,the high-resolution SAR image is transformed by the curvelet tranform.Then then the coefficients of different scales and different directions are chosen to obtain the features through the curvelet inverse transform.Finally,the features are further classified by the deep convolutional ladder network.Curvelet transform has multi-scale and multi-directional characteristics,which makes the extracted features have stronger robustness to noise,thus reducing the noise in the classification results and improving the classification accuracy.The experimental results show that this method can significantly improve the problem of multiple points of misclassification in the classification results and improve the continuity and consistency of the classification areas.3.A high resolution SAR image classification method based on curvelet deep complex convolutional ladder network is proposed.This method changes the operations in the convolutional ladder network to complex valued forms,including complex convolution,complex fully connect,and complex batch normalization operations.The use of the complex form of the convolutional ladder network can fully exploit the amplitude and phase information contained in the complex number,enhance the representation ability of the features,and further improve the classification accuracy.Experimental results can verify that the complex convolutional ladder network improves the classification accuracy of highresolution SAR images,and the noise in the classified images are reduced,and the result is improved significantly.