Weather recognition combining improved ConvNeXt models with knowledge distillation

Table 1. Datasets content description
View table
View in Article
Table 1. Datasets content description
种类描述数量
Sunny 晴天 3 180
Snow 雪天 1 540
Rainy 雨天 1 580
Haze 雾天 1 700

Table 2. Confusion matrix table
View table
View in Article
Table 2. Confusion matrix table
真是情况预测结果（正类）预测结果（负类）
正类 TP（真正例） FN（假反例）
负类 FP（假正例） TN（真反例）

Table 3. Weather classification results of different algorithms for weather-traffic dataset

View table

View in Article

Table 3. Weather classification results of different algorithms for weather-traffic dataset

Model	Accuracy/%	FPS/Hz	FLOPs /G	Params /M
DenseNet^［9］	93.44	33.5	2.88	7.9
ResNet50^［10］	93.5	95.6	4.11	25.56
AlexNet^［11］	93.5	95.6	0.72	61.1
VGG16-TL^［12］	92.81	33.6	15.65	138.36
JVNet^［13］	94.81	11.9	10.11	75.92
Swin transformer^［22］	96.56	45.1	15.14	87.77
ConvNeXt^［15］	96.83	108.9	4.46	28.57
ConvNeXt_F	98.79	111.8	4.46	28.39
MobileNetV3^［23］	83.58	132.7	0.06	2.54
MobileNetV3_Z（ours）	96.22	157.6	0.06	2.54

Table 4. Weather classification results of different algorithms for RSCM2017 dataset

View table

View in Article

Table 4. Weather classification results of different algorithms for RSCM2017 dataset

Model	Accuracy/%	FPS/Hz
DenseNet^［9］	76.87	33.3
ResNet50^［10］	82.63	86.5
AlexNet^［11］	80.82	44.8
VGG16-TL^［12］	80.3	31.6
JVNet^［13］	85.15	81.6
Swin transformer^［22］	87.31	25.6
ConvNeXt^［15］	88.49	99.1
ConvNeXt_F	90.47	96.5
MobileNetV3^［23］	73.27	115.4
MobileNetV3_Z（ours）	84.80	137.6

Table 5. Comparison experiment results of different attention modules
View table
View in Article
Table 5. Comparison experiment results of different attention modules
Model Accuracy（%）
weather-traffic RSCM2017
+SENet 97.14 89.18
+CBMA 97.36 88.28
+SimAm 98.29 90.29

Table 6. Results of teacher model ablation experiment
View table
View in Article
Table 6. Results of teacher model ablation experiment
Model Accuracy/%
weather-traffic RSCM2017
ConvNeXt 96.94 88.49
L_f 98.06 89.29
SimAm 98.29 90.29
L_f+SimAm 98.79 90.47

Table 7. Table of results of experiments at different distillation temperatures
View table
View in Article
Table 7. Table of results of experiments at different distillation temperatures
蒸馏温度T Accuracy（%）
weather-traffic RSCM2017
3 95.19 83.47
5 96.14 84.29
7 96.22 84.80
9 96.22 84.38
11 96.03 84.33
15 96.00 83.80

Table 8. Experimental results of different mixed loss weight parameters
View table
View in Article
Table 8. Experimental results of different mixed loss weight parameters
权重参数α Accuracy（%）
weather-traffic RSCM2017
0.2 95.83 84.07
0.3 96.22 84.80
0.4 95.68 84.31

Tools

Get Citation

Copy Citation Text

Libo LIU, Siyu XI, Zhen DENG. Weather recognition combining improved ConvNeXt models with knowledge distillation[J]. Optics and Precision Engineering, 2023, 31(14): 2123

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites