Reservoir computing based network for few-shot image classification

方法	骨干网络/Attn	Cifar-FS		FC100
方法	骨干网络/Attn	5-way 1-shot	5-way 5-shot	5-way 1-shot	5-way 5-shot
Cp.Nets^［25］	ResNet-12/No	75.40 $\pm$ 0.20	86.80 $\pm$ 0.20	43.80 $\pm$ 0.20	59.70 $\pm$ 0.20
TPMN^［26］	ResNet-12/No	75.50 $\pm$ 0.90	87.20 $\pm$ 0.60	46.93 $\pm$ 0.71	63.26 $\pm$ 0.74
RFS-distill^［27］	ResNet-12/No	73.90 $\pm$ 0.80	86.90 $\pm$ 0.50	44.60 $\pm$ 0.70	60.90 $\pm$ 0.60
MetaOptNet^［28］	ResNet-12/No	72.60 $\pm$ 0.70	84.30 $\pm$ 0.50	41.10 $\pm$ 0.60	55.50 $\pm$ 0.60
MetaQAD^［29］	WRN-28-10/No	75.83 $\pm$ 0.88	88.79 $\pm$ 0.75	-	-
Centroid^［30］	ResNet-18/No	-	-	45.83 $\pm$ 0.48	59.74 $\pm$ 0.56
STANet^［13］	ResNet-12/Yes	74.89 $\pm$ 0.18^*	88.23 $\pm$ 0.11^*	46.27 $\pm$ 0.22^*	62.89 $\pm$ 0.15^*
Main^［14］	ResNet-12/Yes	74.36 $\pm$ 0.45^*	84.13 $\pm$ 0.78^*	44.54 $\pm$ 0.33	58.09 $\pm$ 0.32
Cro-Attention^［15］	ResNet-12/Yes	75.33 $\pm$ 0.14^*	87.94 $\pm$ 0.61^*	45.78 $\pm$ 0.61^*	62.78 $\pm$ 0.66^*
RCFIC	ResNet-12	77.23 $\pm$ 0.32	88.91 $\pm$ 0.19	48.14 $\pm$ 0.41	64.27 $\pm$ 0.83
RCFIC	ResNet-18	79.44 $\pm$ 0.41	89.86 $\pm$ 0.68	50.49 $\pm$ 0.37	66.52 $\pm$ 0.09

Table 2. Classification accuracy on Mini-ImageNet dataset %

View table

View in Article

Table 2. Classification accuracy on Mini-ImageNet dataset %

方法	骨干网络/Attn	5-way 1-shot	5-way 5-shot
DMF^［31］	ResNet-12/No	67.76 $\pm$ 0.46	82.71 $\pm$ 0.31
IEPT^［32］	ResNet-12/No	67.05 $\pm$ 0.44	82.90 $\pm$ 0.30
CTM^［33］	ResNet-18/No	64.12 $\pm$ 0.82	80.51 $\pm$ 0.13
S2M2^［34］	ResNet-18/No	64.06 $\pm$ 0.18	80.58 $\pm$ 0.12
STANet^［13］	ResNet-12/Yes	58.35 $\pm$ 0.57	71.07 $\pm$ 0.39
Main^［14］	ResNet-12/Yes	64.27 $\pm$ 0.35	81.24 $\pm$ 0.26
Cro-Attention^［15］	ResNet-12/Yes	67.19 $\pm$ 0.55	80.64 $\pm$ 0.35
RCFIC	ResNet-12	67.95 $\pm$ 0.57	83.15 $\pm$ 0.33
RCFIC	ResNet-18	69.87 $\pm$ 0.32	84.45 $\pm$ 0.61

Table 3. Cross-domain

View table

View in Article

Table 3. Cross-domain

方法	骨干网络	5-way 1-shot	5-way 5-shot
LFWT^［35］	ResNet-10	47.47 $\pm$ 0.75	66.98 $\pm$ 0.68
LRP^［36］	ResNet-12	46.23 $\pm$ 0.42	66.58 $\pm$ 0.39
S-Shot^［37］	ResNet-18	46.68 $\pm$ 0.49	65.56 $\pm$ 0.70
RCFIC	ResNet-12	48.15 $\pm$ 0.35	67.66 $\pm$ 0.57
RCFIC	ResNet-18	49.24 $\pm$ 0.19	69.07 $\pm$ 0.26

Table 4. Effect of feature enhancement module（taking classification accuracies on Cifar-FS for example）%
View table
View in Article
Table 4. Effect of feature enhancement module（taking classification accuracies on Cifar-FS for example）%
NA OR OA 5-way 1-shot 5-way 5-shot
- - - 73.43 $\pm$ 0.12 84.34 $\pm$ 0.36
- √ - 76.61 $\pm$ 0.57 87.33 $\pm$ 0.72
- - √ 78.66 $\pm$ 0.38 87.62 $\pm$ 0.29
- √ √ 79.44 $\pm$ 0.41 89.86 $\pm$ 0.68

Table 5. Classification accuracy of attention mechanisms generated by different methods on Mini-ImageNet dataset
View table
View in Article
Table 5. Classification accuracy of attention mechanisms generated by different methods on Mini-ImageNet dataset
线性变换卷积储备池 5-way 1-shot 5-way 5-shot
- - - 60.97 $\pm$ 0.22 79.23 $\pm$ 0.07
√ - - 65.27 $\pm$ 0.36 82.33 $\pm$ 0.72
- √ - 63.75 $\pm$ 0.17 81.62 $\pm$ 0.29
- - √ 69.87 $\pm$ 0.32 84.45 $\pm$ 0.61

Table 6. Effect of whether using $β_{1}$ and $β_{2}$
View table
View in Article
Table 6. Effect of whether using $β_{1}$ and $β_{2}$
$β_{2}$ $β_{3}$ 5-way 1-shot 5-way 5-shot
- - 77.12 $\pm$ 0.41 87.37 $\pm$ 0.29
√ √ 79.44 $\pm$ 0.41 89.86 $\pm$ 0.68

Table 7. Effect of different internal topologies of RC

View table

View in Article

Table 7. Effect of different internal topologies of RC

拓扑结构	骨干网络	5-way 1-shot	5-way 5-shot
Random	ResNet-18	67.16 $\pm$ 0.44	80.57 $\pm$ 0.18
Delay line	ResNet-18	65.33 $\pm$ 0.38	78.95 $\pm$ 0.39
Cyclic	ResNet-18	66.97 $\pm$ 0.29	78.52 $\pm$ 0.67
Wigner	ResNet-18	68.44 $\pm$ 0.51	81.38 $\pm$ 0.13
RCFIC	ResNet-18	69.87 $\pm$ 0.32	84.45 $\pm$ 0.61

Tools

Get Citation

Copy Citation Text

Bin WANG, Hai LAN, Hui YU, Jie-long GUO, Xian WEI. Reservoir computing based network for few-shot image classification[J]. Chinese Journal of Liquid Crystals and Displays, 2023, 38(10): 1399

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Research Articles

Received: Dec. 6, 2022

Accepted: --

Published Online: Oct. 25, 2023

The Author Email: Xian WEI (xian.wei@fjirsm.ac.cn)

DOI:10.37188/CJLCD.2022-0407

Topics

Table 1. Classification accuracy on Cifar-FS dataset and FC100 dataset

Table 1. Classification accuracy on Cifar-FS dataset and FC100 dataset

Table 2. Classification accuracy on Mini-ImageNet dataset %

Table 2. Classification accuracy on Mini-ImageNet dataset %

Table 3. Cross-domain

Table 3. Cross-domain

Table 4. Effect of feature enhancement module（taking classification accuracies on Cifar-FS for example）%

Table 4. Effect of feature enhancement module（taking classification accuracies on Cifar-FS for example）%

Table 5. Classification accuracy of attention mechanisms generated by different methods on Mini-ImageNet dataset

Table 5. Classification accuracy of attention mechanisms generated by different methods on Mini-ImageNet dataset

Table 6. Effect of whether using β1 and β2

Table 6. Effect of whether using β1 and β2

Table 7. Effect of different internal topologies of RC

Table 7. Effect of different internal topologies of RC

Table 6. Effect of whether using $β_{1}$ and $β_{2}$

Table 6. Effect of whether using $β_{1}$ and $β_{2}$