Quantitative Spectrometric Analysis Based on a Multi-Branch Atrous Convolutional Network

Guoxi Chen; Yisen Liu; Songbin Zhou; Lulu Zhao

doi:10.3788/LOP213339

Laser & Optoelectronics Progress, Volume. 60, Issue 9, 0930001(2023)

Quantitative Spectrometric Analysis Based on a Multi-Branch Atrous Convolutional Network

Guoxi Chen^1,2, Yisen Liu^2、*, Songbin Zhou², and Lulu Zhao²

¹Faculty of Information Engineering and Automation, Kunming University of Science and Technology, Kunming 650500, Yunnan , China

²Institute of Intelligent Manufacturing, Guangdong Academy of Sciences, Guangzhou 510070, Guangdong , China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(18)

Fig. 1. Spectral curve of one sample for each data set. (a) Tablets (Raman) data set; (b) soil (NIR) data set; (c) wines (NMR) data set

Download full size

Fig. 2. Atrous convolutional layers with a kernel size of 3×1 and different atrous rates

Download full size

Fig. 3. Architectures of proposed atrous convolutional networks. (a) ACCnet; (b) ACPnet

Download full size

Fig. 5. Regression curves of tablets (Raman) data set obtained by ACPnet and contrast methods

Download full size

Fig. 6. Regression curves of soil (NIR) data set obtained by ACPnet and contrast methods

Download full size

Fig. 7. Regression curves of wines (NMR) data set obtained by ACPnet and contrast methods

Download full size

Fig. 8. Regression results obtained by ACCnet with different atrous rates

Download full size

Fig. 9. Regression results obtained by ACPnet with different atrous rates

Download full size

Table 1. Detailed information of three data sets

View table

Table 1. Detailed information of three data sets

Data set	Scan	Sample size	Feature size	Target	Mean	Range	Standard deviation
Soil	NIR	108	1050	Soil organic matter	85.43%	42.91%-95.85%	10.82%
Tablets	Raman	120	3400	Active substance	7.38%	5.12%-8.48%	1.13%
Wines	NMR	40	1500	Ethanol	12.84 g·L^-1	11.19-14.54 g·L^-1	0.78 g·L^-1

Table 2. Detailed parameters for regular CNN, ACCnet, and ACPnet of tablets (Raman) data set

View table

Table 2. Detailed parameters for regular CNN, ACCnet, and ACPnet of tablets (Raman) data set

CNN		ACCnet		ACPnet
Avg-pooling1	3	Avg-pooling1	3	Avg-pooling1			3
Conv1	16×5 PReLU	Conv1 Rate is 2	16×5 PReLU	Conv1_1 Rate is 1	Conv1_2 Rate is 2	Conv1_3 Rate is 5	16×5 PReLU
				Avg-pooling2_1	Avg-pooling2_2	Avg-pooling2_3	3
Avg-pooling2	3	Avg-pooling2	3	Feature fusion			Sum
Conv2	16×5 PReLU	Conv2 Rate is 2	16×5 PReLU	Conv2_1 Rate is 1	Conv2_2 Rate is 2	Conv2_3 Rate is 5	16×5 PReLU
				Avg-pooling3_1	Avg-pooling3_2	Avg-pooling3_3	3
Avg-pooling3	3	Avg-pooling3	3	Feature fusion			Sum
Conv3	16×5 PReLU	Conv3 Rate is 2	16×5 PReLU	Conv3_1 Rate is 1	Conv3_2 Rate is 2	Conv3_3 Rate is 5	16×5 PReLU
				Avg-pooling4_1	Avg-pooling4_2	Avg-pooling4_3	3
Avg-pooling4	3	Avg-pooling4	3	Feature fusion			Sum
Fully-connected	32	Fully-connected	32	Fully-connected			32

Table 3. Detailed parameters for regular CNN, ACCnet, and ACPnet of soil (NIR) data set

View table

Table 3. Detailed parameters for regular CNN, ACCnet, and ACPnet of soil (NIR) data set

CNN		ACCnet		ACPnet
Conv1	32×5 PReLU	Conv1 Rate is 2	32×5 PReLU	Conv1_1 Rate is 1	Conv1_2 Rate is 2	Conv1_3 Rate is 5	32×5 PReLU
				Avg-pooling1_1	Avg-pooling1_2	Avg-pooling1_3	2
Avg-pooling1	2	Avg-pooling1	2	Feature fusion			Sum
Conv2	32×5 PReLU	Conv2 Rate is 2	32×5 PReLU	Conv2_1 Rate is 1	Conv2_2 Rate is 2	Conv2_3 Rate is 5	32×5 PReLU
				Avg-pooling2_1	Avg-pooling2_2	Avg-pooling2_3	2
Avg-pooling2	2	Avg-pooling2	2	Feature fusion			Sum
Conv3	32×5 PReLU	Conv3 Rate is 2	32×5 PReLU	Conv3_1 Rate is 1	Conv3_2 Rate is 2	Conv3_3 Rate is 5	32×5 PReLU
				Avg-pooling3_1	Avg-pooling3_2	Avg-pooling3_3	2
Avg-pooling3	2	Avg-pooling3	2	Feature fusion			Sum

Table 4. Detailed parameters for regular CNN, ACCnet, and ACPnet of wines (NMR) data set

View table

Table 4. Detailed parameters for regular CNN, ACCnet, and ACPnet of wines (NMR) data set

CNN		ACCnet		ACPnet
Avg-pooling1	2	Avg-pooling1	2	Avg-pooling1			2
Conv1	4×5 PReLU	Conv1 Rate is 2	4×5 PReLU	Conv1_1 Rate is 1	Conv1_2 Rate is 2	Conv1_3 Rate is 5	4×5 PReLU
				Avg-pooling2_1	Avg-pooling2_2	Avg-pooling2_3	3
Avg-pooling2	3	Avg-pooling2	3	Feature fusion			Sum
Conv2	4×5 PReLU	Conv2 Rate is 2	4×5 PReLU	Conv2_1 Rate is 1	Conv2_2 Rate is 2	Conv2_3 Rate is 5	4×5 PReLU
				Avg-pooling3_1	Avg-pooling3_2	Avg-pooling3_3	3
Avg-pooling3	3	Avg-pooling3	3	Feature fusion			Sum
Conv3	4×5 PReLU	Conv3 Rate is 2	4×5 PReLU	Conv3_1 Rate is 1	Conv3_2 Rate is 2	Conv3_3 Rate is 5	4×5 PReLU
				Avg-pooling4_1	Avg-pooling4_2	Avg-pooling4_3	2
Avg-pooling4	2	Avg-pooling4	2	Feature fusion			Sum

Table 5. Regression results of ACPnet and contrast methods of tablets (Raman) data set

View table

Table 5. Regression results of ACPnet and contrast methods of tablets (Raman) data set

Method	RMSEC /%	RMSEP /%	RMSECV /%	R²	RPD
PLS	0.30±0.02	0.53±0.10	0.59±0.14	0.75±0.08	2.15±0.38
LS-SVM	0.22±0.06	0.54±0.16	0.57±0.22	0.71±0.18	2.16±0.56
CNN	0.27±0.06	0.44±0.18	0.47±0.17	0.80±0.14	3.12±1.72
ACCnet（rate is 2）	0.29±0.04	0.43±0.16	0.45±0.16	0.80±0.18	2.99±1.17
ACPnet（rate is 1/2/5）	0.23±0.03	0.36±0.20	0.39±0.19	0.85±0.20	3.76±1.40

Table 6. Regression results of ACPnet and contrast methods of soil (NIR) data set

View table

Table 6. Regression results of ACPnet and contrast methods of soil (NIR) data set

Method	RMSEC /%	RMSEP /%	RMSECV /%	R²	RPD
PLS	0.75±0.13	1.67±0.46	2.07±0.57	0.96±0.05	6.18±1.74
LS-SVM	0.24±0.13	1.66±0.71	2.10±0.78	0.95±0.05	5.99±2.33
CNN	1.06±0.09	1.78±0.39	2.16±0.44	0.95±0.04	5.85±2.26
ACCnet（rate is 2）	1.09±0.36	1.67±0.70	1.50±0.65	0.96±0.03	6.45±2.82
ACPnet（rate is 1/2/5）	0.42±0.22	1.27±0.33	1.34±0.34	0.98±0.02	8.26±3.10

Table 7. Regression results of ACPnet and contrast methods of wines (NMR) data set

View table

Table 7. Regression results of ACPnet and contrast methods of wines (NMR) data set

Method	RMSEC /（g·L^-1）	RMSEP /（g·L^-1）	RMSECV /（g·L^-1）	R²	RPD
PLS	0.02±0.02	0.26±0.12	0.43±0.37	0.71±0.29	3.60±2.39
LS-SVM	0.02±0.03	0.29±0.09	0.44±0.30	0.70±0.25	2.52±0.09
CNN	0.18±0.03	0.26±0.11	0.25±0.15	0.68±0.40	3.38±2.11
ACCnet（rate is 2）	0.19±0.02	0.25±0.11	0.15±0.06	0.69±0.49	3.52±1.95
ACPnet（rate is 1/2/5）	0.11±0.01	0.21±0.09	0.14±0.06	0.74±0.39	4.27±2.86

Table 8. Regression results of tablets (Raman) data set obtained by extractor-regressor models
View table
Table 8. Regression results of tablets (Raman) data set obtained by extractor-regressor models
Regressor Feature extractor
CNN ACPnet
RMSECV /% R² RPD RMSECV /% R² RPD
PLS 0.44±0.19 0.80±0.17 3.07±1.32 0.38±0.21 0.83±0.20 3.52±1.45
LS-SVM 0.45±0.21 0.78±0.19 3.10±1.52 0.37±0.25 0.83±0.23 3.99±1.71

Table 9. Regression results of soil (NIR) data set obtained by extractor-regressor models
View table
Table 9. Regression results of soil (NIR) data set obtained by extractor-regressor models
Regressor Feature extractor
CNN ACPnet
RMSECV /% R² RPD RMSECV /% R² RPD
PLS 1.55±0.16 0.96±0.02 7.17±2.56 1.48±0.75 0.96±0.01 7.55±2.50
LS-SVM 1.53±0.62 0.96±0.02 6.82±1.62 1.43±0.56 0.97±0.01 7.34±1.96

Table 10. Regression results of wines (NMR) data set obtained by extractor-regressor models

View table

Table 10. Regression results of wines (NMR) data set obtained by extractor-regressor models

Regressor	Feature extractor
	CNN			ACPnet
	RMSECV /（g·L^-1）	R²	RPD	RMSECV /（g·L^-1）	R²	RPD
PLS	0.25±0.13	0.63±0.54	3.67±1.92	0.22±0.07	0.79±0.16	3.82±2.39
LS-SVM	0.24±0.12	0.67±0.45	3.91±2.48	0.22±0.09	0.82±0.14	3.88±2.43

Tools

Get Citation

Copy Citation Text

Guoxi Chen, Yisen Liu, Songbin Zhou, Lulu Zhao. Quantitative Spectrometric Analysis Based on a Multi-Branch Atrous Convolutional Network[J]. Laser & Optoelectronics Progress, 2023, 60(9): 0930001

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Spectroscopy

Received: Dec. 24, 2021

Accepted: Mar. 3, 2022

Published Online: May. 9, 2023

The Author Email: Liu Yisen (ys.liu@giim.ac.cn)

DOI:10.3788/LOP213339

Topics

laser devices and laser physics

Lasers and Laser Optics

Laser physics

laser manufacturing

Instrumentation, Measurement and Metrology

Table 1. Detailed information of three data sets

Table 1. Detailed information of three data sets

Table 2. Detailed parameters for regular CNN, ACCnet, and ACPnet of tablets (Raman) data set

Table 2. Detailed parameters for regular CNN, ACCnet, and ACPnet of tablets (Raman) data set

Table 3. Detailed parameters for regular CNN, ACCnet, and ACPnet of soil (NIR) data set

Table 3. Detailed parameters for regular CNN, ACCnet, and ACPnet of soil (NIR) data set

Table 4. Detailed parameters for regular CNN, ACCnet, and ACPnet of wines (NMR) data set

Table 4. Detailed parameters for regular CNN, ACCnet, and ACPnet of wines (NMR) data set

Table 5. Regression results of ACPnet and contrast methods of tablets (Raman) data set

Table 5. Regression results of ACPnet and contrast methods of tablets (Raman) data set

Table 6. Regression results of ACPnet and contrast methods of soil (NIR) data set

Table 6. Regression results of ACPnet and contrast methods of soil (NIR) data set

Table 7. Regression results of ACPnet and contrast methods of wines (NMR) data set

Table 7. Regression results of ACPnet and contrast methods of wines (NMR) data set

Table 8. Regression results of tablets (Raman) data set obtained by extractor-regressor models

Table 8. Regression results of tablets (Raman) data set obtained by extractor-regressor models

Table 9. Regression results of soil (NIR) data set obtained by extractor-regressor models

Table 9. Regression results of soil (NIR) data set obtained by extractor-regressor models

Table 10. Regression results of wines (NMR) data set obtained by extractor-regressor models

Table 10. Regression results of wines (NMR) data set obtained by extractor-regressor models