Improved Attention Mechanism MobileNetV2 Network for SERS Classification of Water Pollution

Xueling Li; Jing Yu; Haiyang Zhang; Lu Dong; Zhengdong Zhang; Ke Li; Yaqin Yu; Qi Li

doi:10.3788/LOP242165

Laser & Optoelectronics Progress, Volume. 62, Issue 7, 0730004(2025)

Improved Attention Mechanism MobileNetV2 Network for SERS Classification of Water Pollution

Xueling Li1... Jing Yu1, Haiyang Zhang2, Lu Dong3, Zhengdong Zhang2, Ke Li2, Yaqin Yu2 and Qi Li2,* |Show fewer author(s)

Author Affiliations

¹College of Metrology Measurement and Instrument, China Jiliang University, Hangzhou 310018, Zhejiang , China

²National Institute of Metrology, China, Beijing 100029, China

³Liaoning Inspection, Examination & Certification Centre, Shenyang 110032, Liaoning , China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(12)

Fig. 1. Principle of MP_ECA

Download full size

Fig. 2. Structure of the MobileNetV2_MP_ECA model

Download full size

Fig. 3. Algorithm framework diagram

Download full size

Fig. 4. Spectra and wavelet images of five substances before and after pre-processing. Spectra of (a1) pyrene, (b1) naphthalene, (c1) profenofos, (d1) fenitrothion, and (e1) isocarbophos before and after pre-processing; wavelet images of (a2) pyrene, (b2) naphthalene, (c2) profenofos, (d2) fenitrothion, and (e2) isocarbophos

Download full size

Fig. 5. Iterative accuracies of different models on the validation set

Download full size

Fig. 6. Comparison of heatmaps for raw spectra, pre-processed spectra, and wavelet images. (a) Raw Raman spectra of five substances; (b) heatmaps for the raw Raman spectra of five substances; (c) pre-processed Raman spectra of five substances; (d) heatmaps for the pre-processed Raman spectra of five substances; (e) wavelet images of five substances; (f) heatmaps for the wavelet images of five substances

Download full size

Fig. 7. Confusion matrix

Download full size

Table 1. Structural parameters of the MobileNetV2_MP_ECA model

View table

Table 1. Structural parameters of the MobileNetV2_MP_ECA model

Layer	Configuration ［expansion coefficient， stride］	Input size /（pixel×pixel）	Output channel	Number of layer
Input	—	224×224	3	—
Convolution+BN+ReLU	Conv 3×3， stride is 2	224×224	32	1
InvertedResidual	［1，1］	112×112	16	1
Block-2 layers	［6，2］	112×112	24	2
Block-2 layers	［6，2］	56×56	24	2
Block-2 layers	［6，2］	56×56	32	2
Block-2 layers	［6，2］	28×28	32	2
Block-2 layers	［6，2］	28×28	64	2
Block-2 layers	［6，2］	14×14	64	2
Block-1 layer	［6，1］	14×14	96	1
Block-2 layers	［6，2］	14×14	160	2
Block-2 layers	［6，2］	7×7	160	2
Block-1 layer	［6，1］	7×7	320	1
Convolution+BN+ReLU	Conv 3×3， stride is 1	7×7	1280	1
Global average pooling	Pool H × W	7×7	1280	1
Output	Dropout （0.1）， Linear	1×1	5	—

Table 2. Hyperparameters setting
View table
Table 2. Hyperparameters setting
Hyperparameter Value
Learning rate 0.001
Batch size 64
Number of epoch 50
Optimizer Adam
Loss function Cross-entropy loss

Table 3. Comparison of MobileNetV2 and models with different attention mechanisms

View table

Table 3. Comparison of MobileNetV2 and models with different attention mechanisms

Model	Accuracy /%	P /%	R /%	F1-score /%	Model size /MB	Parameters	FLOPs /MFLOPs
MobileNetV2	95.00	95.64	95.00	95.12	6.08	1550341	228.83
MobileNetV2_MP_ECA	97.83	97.85	97.83	97.83	6.11	1550431	230.20
MobileNetV2_ECA	96.33	96.44	96.33	96.35	6.09	1550371	230.20
MobileNetV2_SE	96.00	96.08	96.00	96.01	7.34	1874629	230.49
MobileNetV2_CBAM	95.67	95.91	95.67	95.71	7.38	1879893	230.32

Table 4. Comparison of different classification models

View table

Table 4. Comparison of different classification models

Data type	Model	Accuracy /%	P /%	R /%	F1-score /%	Model size /MB	Parameters	FLOPs /MFLOPs
Wavelet image	MobileNetV2_MP_ECA	97.83	97.85	97.83	97.83	6.11	1550431	230.20
	VGGNet16	96.17	96.19	96.17	96.16	128.06	134281029	15466.19
	GoogLeNet	95.83	96.10	95.83	95.88	22.30	5840691	1456.95
	ResNet34	95.50	96.97	95.50	95.54	81.36	21287237	3678.23
	DenseNet	94.50	94.56	94.50	95.00	6.64	6958981	2895.99
	EfficientNetV2	94.33	94.48	94.33	94.36	7.12	1866895	412.23
	ShuffleNetV2	93.50	93.98	93.50	93.56	5.01	1258729	151.69
Spectral data	1D-CNN	85.43	84.12	86.72	85.41	1.20	75000	2320.50
Spectral data	SVM	82.37	80.52	83.27	81.88	0.03	25500	—

Table 5. Model classification accuracies by different input images
View table
Table 5. Model classification accuracies by different input images
Input Accuracy
Raw spectrum 93.50
Pre-processed spectrum 95.47
Wavelet image 97.83

Tools

Get Citation

Copy Citation Text

Xueling Li, Jing Yu, Haiyang Zhang, Lu Dong, Zhengdong Zhang, Ke Li, Yaqin Yu, Qi Li. Improved Attention Mechanism MobileNetV2 Network for SERS Classification of Water Pollution[J]. Laser & Optoelectronics Progress, 2025, 62(7): 0730004

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites