基于小波变换和深度网络的着陆地貌图像分类

doi:10.12263/DZXB.20200870

摘要/Abstract

摘要：

针对无人机着陆地貌图像场景复杂、纹理特征丰富等问题，提出一种基于小波变换和深度网络的无人机着陆地貌图像分类算法.利用非下采样小波变换（Non-Subsampled Wavelet Transform，NSWT）的快速压缩能力，将小波变换后的前两层子图系数引入到卷积神经网络（CNN）中，压缩数据量.根据无人机着陆地貌图像的特点，采用轻量化卷积模块设计了15层卷积神经网络.通过支持向量机（SVM）实现复杂地貌场景的正确分类.实验结果表明：所提算法具有良好的特征表达能力，提升了着陆地貌图像的分类准确率.

关键词: 非下采样小波变换, 卷积神经网络, 支持向量机, 图像分类

Abstract:

Aiming at the problems of complex scenes and rich texture features of UAV landing landform image, a classification algorithm of UAV landing landform image based on wavelet transform and deep network is proposed. Using the fast compression capability of non-subsampled wavelet transform (NSWT), the first two layers of sub-image coefficients after wavelet transform are introduced into the convolutional neural network (CNN) to compress the amount of data. According to the characteristics of UAV landing landform image, a 15-layer convolutional neural network is designed using a lightweight convolution module. The correct classification of complex geomorphological scenes is realized by support vector machine (SVM). The experimental results show that the proposed algorithm has good feature expression ability and improves the classification accuracy of landing landform images.

Key words: non-subsampled wavelet transform, convolutional neural network, support vector machine, image classification

中图分类号:

TN911.73

刘芳, 韩笑. 基于小波变换和深度网络的着陆地貌图像分类[J]. 电子学报, 2021, 49(11): 2171-2176.

Fang LIU, Xiao HAN. Image Classification of Landing Landform Based on Wavelet Transform and Deep Network[J]. Acta Electronica Sinica, 2021, 49(11): 2171-2176.

图/表 9

参考文献 15

1	Krizhevsky A, Sutskever II, Hinton G. ImageNet classification with deep convolutional neural networks[A]. Proceedings of the Advances in Neural Information Processing Systems[C]. USA, 2012. 1097 - 1105.
2	方旭, 王光辉, 杨化超,等. 结合均值漂移分割与全卷积神经网络的高分辨遥感影像分类[J]. 激光与光电子学进展, 2018, 55(2):446 - 454.
	Fang X, Wang G H, Yang H C, et al. High resolution remote sensing image classification combining with mean-shift segmentation and fully convolution neural network[J]. Laser & Optoelectronics Progress, 2018, 55 (2):446 - 454.(in Chinese)
3	徐婷婷, 吉晓东, 李文华,等. 基于颜色和纹理特征的胶囊内镜图像分类[J]. 现代电子技术, 2018, 41(19):66 - 70.
	Xu T T, Ji X D, Li W H, et al. Capsule endoscope image classification based on color and texture features [J]. Modern Electronics Technique, 2018, 41 (19): 66 - 70. (in Chinese)
4	张慧娜,李裕梅,傅莺莺.基于Haar-CNN模型的自然场景图像分类的研究[J].四川师范大学学报(自然科学版),2017,40(01):119 - 126.
	Zhang H N, Li Y M, Fu Y Y. Research on natural scene image classification based on Haar-CNN model [J]. Journal of Sichuan Normal University (Natural Science), 2017,40 (01): 119 - 126. (in Chinese)
5	Ioffe S, Szegedy C. Batch normalization: accelerating deep network training by reducing internal covariate shift[J]. Machine Learning, arXiv:,2015.
6	He K, Zhang X, Ren S, et al. Delving deep into rectifiers: surpassing human-level performance on ImageNet classification[A]. IEEE International Conference on Computer Vision (ICCV) [C]. Chile: IEEE, 2015. 1026 - 1034.
7	Clevert D E, Unterthiner T, Hochreiter S. Fast and accurate deep network learning by exponential linear units (elus) [J]. arXiv:2015, 1511.07289.
8	Sainath TN, Mohamed A, Kingsbury B, et al. Deep convolutional neural networks for LVCSR[A]. Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing[C]. Vancouver: IEEE, 2013. 8614 - 8618.
9	李明威. 图像分类中的卷积神经网络方法研究[D].南京邮电大学,2016.
	Li M W. Research of Convolution Neural Network in Image Classification [D]. Nanjing University of Posts and Telecommunications, 2016. (in Chinese)
10	Ju Y, Guo J, Liu S. A deep learning method combined sparse autoencoder with SVM[A]. 2015 International Conference on Cyber-Enabled Distributed Computing and Knowledge Discovery[C]. USA: IEEE, 2015.257 - 260.
11	Zeggada A, Benbraika S, Melgani F, et al. Multilabel conditional random field classification for UAV images[J]. IEEE Geoscience and Remote Sensing Letters, 2018, 15(99):1 - 5.
12	Karen S, Andrew Z, et al. Very deep convolutional networks for large-scale image recognition[J]. Computer Science, arXiv 1409.1556, 2014.
13	He K, Zhang X, Ren S, et al. Deep residual learning for image recognition[A]. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) [C]. USA: IEEE, 2016. 770 - 778.
14	Huang G, Liu Z, Laurens V D M, et al. Densely connected convolutional networks[A]. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) [C]. USA: IEEE, 2017. 2261 - 2269.
15	刘芳, 路丽霞, 黄光伟,等.基于离散余弦变换和深度网络的地貌图像分类[J]. 光学学报, 2018, 38(6):266 - 274.
	Liu F, Lu L X, Huang G W, et al. Landform image classification based on discrete Cosine transformation and deep network [J]. Acta Optica Sinica, 2018, 38 (6): 266 - 274. (in Chinese)

网络层	类型	卷积核	零填充	输出
x	input			256×256
C1	Convolution	5×5	2	256×256×32
	PReLU
P1	Max Pooling	3×3		128×128
C2	Convolution	3×3	1	128×128×64
C3	Convolution	3×3	1	128×128×64
	PReLU
P2	Max Pooling	3×3		64×64
C4	Convolution	3×3	1	32×32×64
C5	Convolution	3×3	1	32×32×64
C6	Convolution	1×1		32×32×32
	PReLU
P3	Max Pooling	3×3		16×16
C7	Convolution	3×3		14×14×32
C8	Convolution	3×3		14×14×32
C9	Convolution	3×3		14×14×32
	PReLU
P4	Global Mean Pooling	3×3		7×7
C10	Convolution	7×7		1×1×32
C11	Convolution	1×1		1×1×C
	SVM			v

网络层	类型	卷积核	零填充	输出
x	input			256×256
C1	Convolution	5×5	2	256×256×32
	PReLU
P1	Max Pooling	3×3		128×128
C2	Convolution	3×3	1	128×128×64
C3	Convolution	3×3	1	128×128×64
	PReLU
P2	Max Pooling	3×3		64×64
C4	Convolution	3×3	1	32×32×64
C5	Convolution	3×3	1	32×32×64
C6	Convolution	1×1		32×32×32
	PReLU
P3	Max Pooling	3×3		16×16
C7	Convolution	3×3		14×14×32
C8	Convolution	3×3		14×14×32
C9	Convolution	3×3		14×14×32
	PReLU
P4	Global Mean Pooling	3×3		7×7
C10	Convolution	7×7		1×1×32
C11	Convolution	1×1		1×1×C
	SVM			v

准确率	Dataset1	Dataset2	Database
Method 1	89.73%	88.61%	88.75%
Method 2	97.49%	97.02%	97.38%
WT-CNN	97.37%	96.85%	97.25%

准确率	Dataset1	Dataset2	Database
Method 1	89.73%	88.61%	88.75%
Method 2	97.49%	97.02%	97.38%
WT-CNN	97.37%	96.85%	97.25%

推理时间（s）	Dataset1	Dataset2	Database
Method 1	0.033	0.032	0.031
Method 2	0.068	0.061	0.063
WT-CNN	0.036	0.035	0.034