Researching | Lightweight deep global-local knowledge distillation network for hyperspectral image scene classification

Optics and Precision Engineering, Volume. 31, Issue 17, 2598(2023)

Lightweight deep global-local knowledge distillation network for hyperspectral image scene classification

Yingxu LIU1... Chunyu PU1, Diankun XU2, Yichuan YANG1 and Hong HUANG1,* |Show fewer author(s)

Author Affiliations

¹Key Laboratory of Optoelectronic Technology and Systems of the Education Ministry of China， Chongqing University， Chongqing400044， China

²Measurement and Control Technology and Instrument major， College of Optoelectronic Engineering， Chongqing University， Chongqing400044， China

show less

Abstract Get PDF(in Chinese)

Figures&Tables (13)

References (24)

Paper Information

Figures & Tables(13)

Fig. 1. Framework of proposed Lightweight Deep Globle-local Knowledge Distillation（LDGLKD） network

Download full size

View in Article

Fig. 2. Detail structure of layer l in ViT model

Download full size

View in Article

Fig. 3. Overall build process for OHID-SC dataset

Download full size

View in Article

Fig. 4. Examples of scene in constructed OHID-SC dataset

Download full size

View in Article

Fig. 5. Examples and numbers of scene in HSRS-SC dataset

Download full size

View in Article

Fig. 6. OAs with respect to different temperatures

Download full size

View in Article

Fig. 7. Effect of KD coefficient α on OAs

Download full size

View in Article

Fig. 8. Ablation experiment results

Download full size

View in Article

Fig. 9. Confusion matrix results of LDGLKD network

Download full size

View in Article

Fig. 10. OAs with different training data percentages for comparison methods

Download full size

View in Article

Table 1. Steps of LDGLKD algorithm

View in Article

Table 1. Steps of LDGLKD algorithm

算法：轻量化深度全局-局部知识蒸馏

训练过程：

阶段1：以小的知识蒸馏系数 $α$ 进行训练

输入：训练集图像 $I_{i n}$ ，温度 $T$ ，知识蒸馏系数 $α$ ，教师模型ViT，学生模型VGG16，教师模型函数集合 $f_{t e a} ()$ ，学生模型函数集合 $f_{s t u} ()$ ；

输出：经过优化的教师模型与学生模型；

1： $F_{t e a} \leftarrow f_{t e a} (I_{i n})$ ；

2： $F_{s t u} \leftarrow f_{s t u} (I_{i n})$ ；

3： $L_{K D} \leftarrow T^{2} \times K L d i v (Q_{s t u}^{T}, Q_{t e a}^{T})$ ， $Q_{s t u}^{T}$ 和 $Q_{t e a}^{T}$ 由公式（18）和（19）计算；

4： $L_{f i r} \leftarrow α L_{K D} + (1 - α) (L_{C}^{t e a} + L_{C}^{s t u})$ ；

5：以 $L_{f i r}$ 为损失函数对教师模型以及学生模型进行共同优化；

阶段2：以大的知识蒸馏系数

α

进行训练

输入：训练集图像

I_{i n}

，温度

T

，知识蒸馏系数

α

，教师模型ViT，学生模型VGG16，教师模型函数集合

f_{t e a} ()

，学生模型函数集合

f_{s t u} ()

；

输出：经过优化的学生模型；

6： $F_{t e a} \leftarrow f_{t e a} (I_{i n})$ ；

7： $F_{s t u} \leftarrow f_{s t u} (I_{i n})$ ；

8： $L_{K D} \leftarrow T^{2} \times K L d i v (Q_{s t u}^{T}, Q_{t e a}^{T})$ ； $Q_{s t u}^{T}$ 和 $Q_{t e a}^{T}$ 由公式（18）和（19）计算；

9： $L_{s e c} \leftarrow 2 α L_{K D} + (1 - 2 α) L_{C}^{s t u}$ ；

10：以 $I_{s e c}$ 为损失函数对学生模型进行优化，并停止对教师模型的训练；

测试阶段：

场景分类

输入：测试集图像， $I_{i n}^{t e s t}$ ，学生模型VGG16，学生模型函数集合， $f_{s t u} ()$ ；

输出：预测标签 $Y_{l a b e l}$ ；

11. $F_{s t u} \leftarrow f_{s t u} (I_{i n}^{t e s t})$ ；

12. $Y_{l a b e l} \leftarrow s o f t m a x (F_{s t u})$ ；

13.生成每一张测试集图像的语义标签

Y_{l a b e l}

。

Table 2. Results of comparison algorithm ［OA±STD］

View in Article

Table 2. Results of comparison algorithm ［OA±STD］

Method	OHID-SC	HSRS-SC
ResNet101	75.86±1.41	95.95±0.26
ResNet18	80.34±0.36	96.43±0.47
GoogleNet	58.17±3.80	87.81±0.04
EffcientNet	83.81±0.66	97.21±0.64
VGG16	81.78±0.50	95.13±0.22
SKAL-R	55.14±0.42	96.08±0.15
SKAL-V	59.80±0.43	96.01±0.24
ACRNet-R	60.39±0.18	95.75±0.07
ACRNet-M	65.02±0.14	97.96±0.41
LDGLKD	91.62±0.20	97.96±0.04

Table 3. Running time of comparison algorithms

View in Article

Table 3. Running time of comparison algorithms

Method	Dataset	OHID-SC		HSRS-SC
ResNet101	Train	9	52
ResNet101	Test	45	126
ResNet18	Train	8	49
ResNet18	Test	19	121
GoogleNet	Train	9	50
GoogleNet	Test	27	123
EfficientNet	Train	10	51
EfficientNet	Test	56	125
VGG16	Train	7	52
VGG16	Test	15	129
SKAL-R	Train	5	11
SKAL-R	Test	14	34
SKAL-V	Train	5	15
SKAL-V	Test	15	36
ACRNet-R	Train	2	4
ACRNet-R	Test	5	13
ACRNet-M	Train	2	5
ACRNet-M	Test	5	14
LDGLKD	Train	8	43
LDGLKD	Test	13	108

Tools

Get Citation

Copy Citation Text

Yingxu LIU, Chunyu PU, Diankun XU, Yichuan YANG, Hong HUANG. Lightweight deep global-local knowledge distillation network for hyperspectral image scene classification[J]. Optics and Precision Engineering, 2023, 31(17): 2598

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Information Sciences

Received: Jan. 3, 2023

Accepted: --

Published Online: Oct. 9, 2023

The Author Email: HUANG Hong (hhuang@cqu.edu.cn)

DOI:10.37188/OPE.20233117.2598

Topics

laser devices and laser physics

Lasers and Laser Optics

laser manufacturing

Instrumentation, Measurement and Metrology