
Lightweight Fully-Connected Tensorial Mapping Network for Hyperspectral Image Classification
LIN Zhi-xin, ZHENG Yu-bang, MA Tian-yu, WANG Rui, LI Heng-chao
ACTA ELECTRONICA SINICA ›› 2024, Vol. 52 ›› Issue (10) : 3541-3551.
Lightweight Fully-Connected Tensorial Mapping Network for Hyperspectral Image Classification
In recent years, convolutional neural networks have demonstrated outstanding performance in HSIC (Hyperspectral Image Classification). However, the improvement of model performance involves adopting deeper and broader network architectures, leading to an increased number of parameters and operations, thus hindering deployment in airborne or on-board devices. To this end, this paper introduces a HSIC method based on the LiteFCTMN (Lightweight Fully-Connected Tensorial Mapping Network). We design two convolutional units based on the mapping way of FCTN (Fully-Connected Tensor Network) decomposition and the structural characteristics of HSIs. By mapping the original convolution kernel to multiple small-sized convolution kernels with fully-connected structures, the complexity of the novel units is reduced while their expressiveness is improved. In addition, the RDT (Residual Double-Branch Tensorial) module is constructed using the designed units. In this module, two branches share the same weights, and a channel split operation is employed to reduce the number of feature channels, thereby reducing complexity. The proposed model strategically leverages both local spatial-spectral information from RDT and global spectral information from the new units, resulting in enhanced classification performance and reduced hardware consumption. Experimental results on three widely used HSI datasets demonstrate that the proposed model achieves superior classification performance and lower complexity compared to the state-of-the-art works.
hyperspectral image classification / model compression / fully-connected tensor network decomposition / convolutional neural network / tensorial neural network / lightweight convolutional module {{custom_keyword}} /
表1 Conv3D和FCTNMConv的空间和计算杂度比较 |
单元 | 空间复杂度 | 计算复杂度 |
---|---|---|
Conv3D | 15 552 | 122 192 064 |
FCTNMConv3D4 | 576 | 6 788 448 |
FCTNMConv3D3 | 348 | 3 582 792 |
表2 LiteFCTMN模型在Indian Pines数据集的参数设置 |
层名称 | 卷积核尺寸 | 输出尺寸 |
---|---|---|
输入层 | — | |
Conv3D | | |
通道分割 | — | |
PW Conv3D | | |
FCTNMConv3D3 | | |
| | |
| | |
PWConv3D | | |
FCTNMConv3D3 | | |
| | |
| | |
拼接 | — | |
残差连接 | — | |
FCTNMConv3D4 | | |
| | |
| | |
| | |
GAP | — | |
FC | — | |
算法1 LiteFCTMN构建流程 | ||
输入: (1)高光谱图像训练集 (2)LiteFCTMN模型参数 (3)训练轮次 输出:高光谱图像测试集样本预测类别 1:随机初始化参数 2:for 3: 通过标准三维卷积计算 4: 通过式( 5: 通过 6: 计算交叉熵损失 7: 通过随机梯度下降更新 8:end for 9:将测试集样本 |
表3 三个数据集训练样本和测试样本个数 |
Indian Pines | Kennedy Space Center | Houston | ||||||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|
序号 | 颜色 | 类别 | 训练 | 测试 | 颜色 | 类别 | 训练 | 测试 | 颜色 | 类别 | 训练 | 测试 |
1 | Alfalfa | 5 | 41 | Srub | 23 | 738 | Health grass | 198 | 1 053 | |||
2 | Corn-notill | 71 | 1 357 | CP swamp | 7 | 236 | Stressed grass | 190 | 1 064 | |||
3 | Corn-mintill | 41 | 789 | CP hammock | 8 | 248 | Synthetic grass | 192 | 505 | |||
4 | Corn | 12 | 225 | Slash pine | 8 | 244 | Trees | 188 | 1 056 | |||
5 | Grass-pasture | 24 | 459 | Oak/Broadleaf | 5 | 156 | Soil | 186 | 1 056 | |||
6 | Grass-trees | 37 | 693 | Hardwood | 7 | 222 | Water | 182 | 143 | |||
7 | Grass-pasture-mowed | 5 | 23 | Swamp | 3 | 102 | Residential | 196 | 1 072 | |||
8 | Hay-windrowed | 24 | 454 | Graminoid | 13 | 418 | Commercial | 191 | 1 053 | |||
9 | Oats | 5 | 15 | Spartina marsh | 16 | 504 | Road | 193 | 1 059 | |||
10 | Soybean-notill | 49 | 923 | Cattail marshl | 49 | 923 | Highway | 191 | 1 036 | |||
11 | Soybean-mintill | 109 | 2 346 | Salt marsh | 12 | 392 | Railway | 181 | 1 054 | |||
12 | Soybean-clean | 30 | 563 | Mud flats | 13 | 406 | Parking Lot 1 | 192 | 1 041 | |||
13 | Wheat | 12 | 193 | Water | 15 | 488 | Parking Lot 2 | 184 | 285 | |||
14 | Woods | 63 | 1 202 | Tennis court | 181 | 247 | ||||||
15 | Buildings-Grass-Trees-Drives | 19 | 369 | Running track | 187 | 473 | ||||||
16 | Stone-Steel-Towers | 6 | 87 | |||||||||
总计 | 512 | 9 737 | 158 | 5 053 | 2 832 | 12 197 |
表4 不同模型在三个数据集上的分类性能 (%) |
模型 | Indian Pines | Kennedy Space Center | Houston | ||||||
---|---|---|---|---|---|---|---|---|---|
OA | AA | | OA | AA | | OA | AA | | |
SVM | 79.66 | 79.09 | 76.68 | 80.44 | 71.08 | 78.13 | 70.85 | 68.35 | 68.43 |
3DCNN | 95.69 | 95.53 | 95.09 | 76.70 | 67.79 | 73.94 | 83.51 | 83.38 | 82.12 |
SSRN | 94.50 | 95.35 | 93.73 | 94.32 | 90.12 | 93.67 | 84.09 | 85.70 | 82.77 |
DBDA | 94.37 | 95.06 | 93.58 | 94.74 | 91.30 | 94.15 | 84.93 | 87.21 | 83.66 |
LiteDenseNet | 95.57 | 95.89 | 94.95 | 96.27 | 93.14 | 95.84 | 85.47 | 87.68 | 84.27 |
LiteDepthwiseNet | 95.95 | 96.67 | 95.38 | 96.30 | 93.49 | 95.89 | 85.11 | 87.07 | 83.89 |
LiteFCTMN | 96.51 | 96.93 | 96.02 | 96.92 | 94.34 | 96.58 | 86.49 | 88.02 | 85.35 |
表5 不同模型在三个数据集上的复杂度 |
模型 | Indian Pines | Kennedy Space Cente | Houston | |||
---|---|---|---|---|---|---|
Params/k | FLOPs/M | Params/k | FLOPs/M | Params/k | FLOPs/M | |
3DCNN | 1 803.80 | 273.20 | 1 802.90 | 273.20 | 1 803.50 | 273.20 |
SSRN | 364.17 | 95.58 | 327.23 | 83.91 | 278.13 | 68.35 |
DBDA | 382.35 | 108.17 | 338.21 | 94.99 | 280.08 | 77.41 |
LiteDenseNet | 852.30 | 171.81 | 748.44 | 150.56 | 610.32 | 122.22 |
LiteDepthwiseNet | 51.71 | 49.00 | 46.35 | 42.97 | 39.56 | 34.93 |
LiteFCTMN | 3.68 | 17.87 | 3.41 | 15.67 | 3.47 | 12.74 |
表6 不同张量分解模型分类性能与复杂度比较 |
模型 | Indian Pines | Kennedy Space Center | Houston | ||||||
---|---|---|---|---|---|---|---|---|---|
OA/% | Params/k | FLOPs/M | OA/% | Params/k | FLOPs/M | OA/% | Params/k | FLOPs/M | |
CP-6 | 96.28 | 3.48 | 18.30 | 95.84 | 3.22 | 16.05 | 84.45 | 3.66 | 13.04 |
Tucker-3 | 93.73 | 3.03 | 18.09 | 94.33 | 2.81 | 15.86 | 83.71 | 3.30 | 12.88 |
TT-4 | 95.82 | 4.52 | 20.36 | 95.39 | 4.15 | 17.85 | 83.90 | 4.65 | 14.50 |
LiteFCTMN | 96.51 | 3.68 | 17.87 | 96.92 | 3.41 | 15.67 | 86.49 | 3.47 | 12.74 |
1 |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
2 |
赵旭东, 陶然, 李伟, 等. 高光谱图像分数域信息提取理论与方法进展[J]. 电子学报, 2022, 50(12): 2874-2883.
{{custom_citation.content}}
{{custom_citation.annotation}}
|
3 |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
4 |
方健, 杨劲翔, 肖亮. 基于多层小波深度聚合网络的高光谱图像超分辨率方法[J]. 电子学报, 2024, 52(1): 201-216.
{{custom_citation.content}}
{{custom_citation.annotation}}
|
5 |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
6 |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
7 |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
8 |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
9 |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
10 |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
11 |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
12 |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
13 |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
14 |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
15 |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
16 |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
17 |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
18 |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
19 |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
20 |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
21 |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
22 |
Grupo de Inteligencia Computacional de la Universidad del PAÍS Vasco. Hyperspectral remote sensing scenes[EB/OL]. (2021-07-12)[2024-05-08].
{{custom_citation.content}}
{{custom_citation.annotation}}
|
23 |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
24 |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
25 |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
26 |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
27 |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
{{custom_ref.label}} |
{{custom_citation.content}}
{{custom_citation.annotation}}
|
/
〈 |
|
〉 |