摘要： 深度神经网络广泛应用于遥感图像场景分类任务中并能大幅提高分类精度，但隐藏层数较少的神经网络在标记数据不足的遥感场景分类中泛化能力较低，而隐层较多的网络往往需要较大的计算量和模型存储空间，限制了其在嵌入式设备上的应用。提出一种针对遥感图像场景分类的多粒度特征蒸馏方法，将深度网络不同阶段的特征与最终的类别概率同时作为浅层模型的监督信号，使得浅层模型能够同时学习高级与低级的语义特征，从而提高浅层模型的分类性能与泛化能力。在UC Merced Land-Use和SIRI-WHU 2个数据集上的实验结果表明，该方法能使模型在大幅降低网络参数量和计算量的情况下明显提高分类性能，与传统知识蒸馏方法相比，其分类精度更高。

关键词: 遥感图像, 多粒度特征蒸馏, 卷积神经网络, 模型压缩, 深度学习

Abstract: Deep neural networks have been widely used in remote sensing images classification and significantly improve the classification accuracy.However,the networks with fewer hidden layers lack generalization ability when classifying remote sensing scenes with insufficient labeled data while the networks with more hidden layers require a large number of computation and storage space resources,which prevents them from further deployment on low-end embedded devices.To solve the problem,this paper proposes a multi-granularity feature distillation method for the scene classification of remote sensing images.The method takes the features of different stages of deep networks and the ultimate category probabilities as the supervision signal of the shallow network.So the shallow model can learn the high-level semantic features as well as the low-level features at the same time,enhancing the classification performance and generalization ability of the shallow model. Experimental results on the UC Merced Land-Use and SIRI-WHU datasets demonstrate that the proposed method can significantly improve the classification performance of the shallow model while reducing the network parameters and computing resources greatly.In addition,the proposed method outperforms the existing knowledge distillation methods in terms of classification accuracy.

Key words: remote sensing image, multi-granularity feature distillation, Convolutional Neural Networks(CNN), model compression, deep learning

中图分类号: 

• TP391