基于泛化空间正则自动编码器的遥感图像识别
Remote sensing image recognition based on generalized regularized auto-encoders

为了解决遥感图像中的合成孔径雷达（synthetic aperture radar，SAR）图像的样本稀缺问题，该文提出了针对这一小样本问题的泛化空间和泛化样本理论，将机器学习的分类问题转化为泛化空间中的样本密度估计问题。首先，通过研究泛化空间方法，针对有限样本的识别分类问题建立了样本密度估计模型，并从理论上验证了泛化空间方法的可行性；其次，在正则化自动编码器模型中，加入了泛化规则作为新的正则化因子对图像重构误差进行约束，针对有限样本问题建立泛化正则自动编码器（generalized auto-encoders，GAE），并提出利用该算法进行图像识别的模型；最后，将该模型应用于遥感图像小样本目标识别问题中。实验结果表明：GAE在SAR图像中具有最优的小样本学习能力，在样本数量有限的情况下，该方法表现出最小的重构误差和测试错误率。在小样本输入情况下，GAE模型实现了对MSTAR图像以及船舶SAR图像的识别分类，进一步证明了该算法相比于同类算法在SAR图像小样本识别问题中更具有优势。
Abstract：This paper addresses the problem of sample scarcity in SAR (synthetic aperture radar) images, which are very useful remote sensing sources but are very costly to obtain. SAR images have limited samples which makes it hard to get a standardized “good” model for even simple classification tasks. Then, the limited samples are expanded via generalization of the sample space using prior and a priori information including labels. Then, auto-learning models, regularized auto-encoders, are used to extract useful features from the generalized sample space. A regularization term is used to penalize generalized samples in an image recognition algorithm. The method produces good results on limit-sized subsets from the MNIST database and the SAR MSTAR image dataset. With limited input samples, the method gives better reconstruction errors and recognition accuracies in the SAR ship images than the state-of-art methods. Thus, this algorithm is shown to be successful in both theoretical analyses and tests.