基于剪枝网络的知识蒸馏对遥感卫星图像分类方法

针对目前遥感图像在应用卷积神经网络分类时需要大量计算,并占用大量内存的问题,提出了一种基于剪枝网络的知识蒸馏对遥感图像分类方法.以模型剪枝理论为基础,在网络结构中引入注意力机制,加强对重要特征的提取之后,并对网络进行模型剪枝,然后引入知识蒸馏技术对模型进行迁移学习,补偿模型剪枝之后分类精度的损失.为了证明方法的先进性与可靠性,利用在NWPU-RESISC45遥感卫星数据集上,与同类算法进行对比实验.实验结果表明,所提方法不仅在分类精度有更好的表现,并且在模型大小上更具有优势.

Knowledge distillation method for remote sensing satellite image classification based on pruning network

At present, the application of convolutional neural network in remote sensing image classification requires a lot of computation and occupies a lot of memory. This paper proposed a knowledge distillation method based on pruning network to classify remote sensing images. Based on the model pruning theory, it introduced attention mechanism into the network structure to strengthen the extraction of important features, and carried out model pruning on the network. Then it introduced knowledge distillation technology to carry out transfer learning to compensate the loss of classification accuracy after pruning. In order to prove the advance and reliability of the method, it carried out a comparison experiment with similar algorithms on NWPU-RESISC45 remote sensing satellite data set. The experimental results show that the proposed method not only has better performance in classification accuracy, but also has more advantages in model size.