基于知识自蒸馏的轻量化复杂遥感图像精细分类方法

Lightweight Fine Classification of Complex Remote Sensing Images Based on Self Knowledge Distillation

基于遥感图像开展地物要素分类和提取,是构建数字化战场环境的技术基础,面临着场景多变、类别多样、噪声干扰等挑战.经典深度学习模型结构复杂、计算量大,难以满足低性能、低功耗边缘计算环境下的实时信息处理要求.提出了一种基于知识自蒸馏的轻量化复杂遥感图像精细分类方法,通过构建一步式自蒸馏框架,实现网络从高层到低层的知识迁移.同时通过金字塔池化融合不同尺度的特征信息,显著提升全局上下文信息的利用率,解决轻量化分类模型由于参数较少、复杂度低,导致场景精细分类精度低的难题.在Vaihingen等公开遥感数据集和我国高分遥感数据集上开展实验,较国际同类方法具有更优的分类性能.

The classification and extraction of ground features based on remote sensing images is the technical basis of building a digital battlefield environment.However,it is faced with challenges such as changeable scenes,diverse categories and noise interference.Traditional deep learning model has complex structure and large amount of computation,which is difficult to meet the requirements of real-time information processing in low-performance,low-power edge computing environment.In this paper,a lightweight and complex remote sensing image fine classification method based on self knowledge distillation is proposed.The method is based on a one-step self knowledge distillation framework to improve the performance of the lightweight model by transferring knowledge from deeper portion of the networks to shallow ones.At the same time,pyramid pooling is used to fuse feature information of different scales,which significantly improves the utilization rate of global context information.This method solves the problem of low precision of remote sensing scene fine classification due to less parameters and low complexity of lightweight classification model.The proposed method achieved state-of-the-art performance and yielded a new record of pixel accuracy 80.76%on Vaihingen and other public remote sensing datasets.