面向遥感影像场景分类的类中心知识蒸馏方法

Class-centric Knowledge Distillation for RSI Scene Classification

卷积神经网络已广泛应用于遥感影像场景分类任务,然而优秀的模型体量大,无法部署到资源受限的边缘设备中,直接应用现有的知识蒸馏方法压缩模型,忽略了场景数据的类内多样性和类间相似性。为此,本文提出一种类中心知识蒸馏方法,旨在获得一个紧凑高效且精度高的遥感影像场景分类网络。首先对预训练的教师网络进行微调,然后基于设计的类中心蒸馏损失将教师网络强大的特征提取能力迁移到学生网络,通过约束师生网络提取的同类特征分布中心的距离完成知识的转移,同时在蒸馏过程中结合真值标签训练,最后学生网络单独用于预测。实验在4个数据集上与8种先进的蒸馏方法在不同训练比率、不同师生架构下进行了比较,本文方法均达到最高分类精度。其中,在训练比率为60%的RSC11、UCM、RSSCN7及AID数据集中,相比于性能最好的其他蒸馏方法,师生网络属同系列时分类总体精度分别提升了2.42%、2.74%、2.95%和1.07%。相似技术对比实验及可视化分析进一步证明了本文方法优异的性能。本文所提出的类中心知识蒸馏方法更好地传递了复杂网络所提取的类内紧凑、类间离散的特征知识,提高了轻量网络分类的性能。

Convolutional neural networks have been widely used in the task of Remote Sensing Image Scene Classification(RSISC)and have achieved extraordinary performance.However,these excellent models have large volume and high computational cost,which cannot be deployed to resource-constrained edge devices.Moreover,in the RSISC task,the existing knowledge distillation method is directly applied to the compression model,ignoring the intra-class diversity and inter-class similarity of scene data.To this end,we propose a novel class-centric knowledge distillation method,which aims to obtain a compact,efficient,and accurate network model for RSISC.The proposed class-centric knowledge distillation framework for remote sensing image scene classification consists of two streams,teacher network flow and student network flow.Firstly,the remote sensing image scene classification dataset is sent into the teacher network pre-trained on a large-scale dataset to fine-tune the parameters.Then,the class-centric knowledge of the hidden layer is extracted from the adjusted teacher network and transferred to the student network based on the designed class center distillation loss,which is realized by constraining the distance of the distribution center of similar features extracted by the teacher and student network,so that the student network can learn the powerful feature extraction ability of the teacher network.The distillation process is combined with the truth tag supervision.Finally,the trained student network is used for scene prediction from remote sensing images alone.To evaluate the proposed method,we design a comparison experiment with eight advanced distillation methods on classical remote sensing image scene classification with different training ratios and different teacher-student architectures.Our results show that:compared to the best performance of other distillation methods,in the case of the teacher-student network belonging to the same series,the overall classification accuracy of our proposed method is increased by 1.429%and 2.74%,respectively,with a given training ratio of 80%and 60%;and in the case of teacher-student networks belonging to different series,the classification accuracy is increased by 0.238%and 0.476%,respectively,with the two given ratios.Additionally,supplementary experiments are also carried out on a small data set of RSC11 with few classes and few samples,a multi-scale data set of RSSCN7 with few classes and multiple books,and a large complex data set of AID with many classes of heterogeneous samples.The results show that the proposed method has good generalization ability.Trough the comparison experiments with similar techniques,it is found that the proposed method can maintain excellent performance in challenging categories through confusion matrix,and the proposed distillation loss function can better deal with noise through testing error curve.And visualization analysis also shows that the proposed method can effectively deal with the problems of intra-class diversity and inter-class similarity in remote sensing image scenes.