| 引用本文: | 刘向举,吴文彦,蒋社想.面向遥感图像场景分类的轻量级沙漏密集网络(J/M/D/N,J:杂志，M：书，D：论文，N：报纸).期刊名称,2025，42（4）：17-26 |
| CHEN X. Adap tive slidingmode contr ol for discrete2ti me multi2inputmulti2 out put systems[ J ]. Aut omatica, 2006, 42(6): 4272-435 |
|
| 摘要: |
| 目的 在具有复杂空间结构和地理布局的遥感图像场景分类任务中,深度卷积神经网络( CNNs) 虽然具有更
好的分类性能,但是通常具有高复杂度,存在不适用于移动或嵌入式设备等问题,针对此,提出一种新的轻量级沙
漏密集网络( LHD-NET) ,以实现分类精度和模型复杂性之间的良好平衡。 方法 首先通过具有特征补偿机制的浅
层混合下采样结构提取浅层信息,在保证信息充分提取的同时减少后续层的参数数量,从而在保持模型轻量级的
同时提高性能;然后在沙漏结构间采用密集连接以提高特征复用,在一定程度上避免了梯度消失,促进了信息传
递;最后利用一个具有较高语义信息的卷积层特征来指导多层特征聚合,以此来提高分类器的性能,同时训练过程
中采用基于标签平滑的交叉熵损失函数,对真实标签进行平滑处理,相比于普通交叉熵损失函数能够有效提高鲁
棒性和减轻模型过拟合问题。 结果 实 验 结 果 表 明: 该 模 型 在 5. 4 M 参 数 量 下 取 得 了 显 著 的 分 类 性 能, 在 UC
Merced Land-Use、SIRI-WHU 和 NWPU-RESISC45 3 个公开遥感数据集上分别取得了 99. 19%、97. 75%和 92. 38%
的平均分类准确率。 结论 通过实验结果可证明所提模型能够在少量参数下便取得较好分类性能,相较于深度神经网
络在保持高分类精度的前提下能显著降低模型参数量,对遥感图像场景分类任务及模型轻量化具有一定的参考价值。
关键词:遥感图像;场景分类;轻量级;卷积神经网络 |
| 关键词: 遥感图像 场景分类 轻量级 卷积神经网络 |
| DOI： |
| 分类号: |
| 基金项目: |
|
| Lightweight Hourglass Dense Network for Remote Sensing Image Scene Classification |
|
LIU Xiangju WU Wenyan JIANG Shexiang
|
|
School of Computer Science and Engineering Anhui University of Science and Technology Anhui Huainan 232001
China
|
| Abstract: |
| Objective In the task of remote sensing image scene classification with complex spatial structures and
geographical layouts although deep convolutional neural networks CNNs have better classification performance they
often have high complexity and are not suitable for mobile or embedded devices. Therefore a new lightweight hourglass
dense network LHD-NET was proposed to achieve a good balance between classification accuracy and model complexity.
Methods Firstly shallow information was extracted through a shallow mixed downsampling structure with a feature
compensation mechanism. This structure can reduce the number of parameters in subsequent layers while ensuring
sufficient information extraction thus improving performance while keeping the model lightweight. Then dense
connections were used between hourglass structures to improve feature reuse and to some extent avoid gradient
disappearance which promoted information transfer. Finally a convolutional layer feature with high semantic information
was used to guide multi-layer feature aggregation so as to improve the performance of the classifier. Meanwhile during
the training process the cross-entropy loss function based on label smoothing was employed to smooth the true labels, which could effectively improve robustness and alleviate overfitting compared with ordinary cross-entropy loss functions.
Results Experimental results showed that the model achieved significant classification performance with only 5. 4 M
parameters obtaining average classification accuracies of 99. 19% 97. 75% and 92. 38% on three publicly available
remote sensing datasets namely UC Merced Land-Use SIRI-WHU and NWPU-RESISC45 respectively.
Conclusion Experimental results demonstrate that the proposed model can achieve good classification performance with a
small number of parameters. Compared with deep neural networks the proposed model significantly reduces the number of
model parameters while maintaining high classification accuracy providing certain reference values for remote sensing
image scene classification tasks and model lightweighting. |
| Key words: remote sensing image scene classification lightweight convolutional neural network |
