Few-Shot Image Classification Algorithm of Graph Neural Network Based on Swin Transformer

Table 1. Information in experimental dataset
View table
Table 1. Information in experimental dataset
Dataset Number of images Number of classes Number of Train/Val/Test samples
CUB-200-2011 11788 200 100/50/50
Stanford Dogs 20580 120 70/20/30
Stanford Cars 16185 196 130/17/49
Mini-Imagenet 60000 100 64/16/20

Table 2. Classification accuray on CUB-200-2011 dataset unit: %

View table

Table 2. Classification accuray on CUB-200-2011 dataset unit: %

Method	Backbone	5-way 1-shot	5-way 5-shot
ProtoNet^［6］	Conv-4	64.82±0.23	85.74±0.14
RelationNet^［8］	Conv-4	63.94±0.92	77.87±0.64
DN4^［9］	Conv-4	57.45±0.80	84.41±0.58
FAN^［11］	Conv-4	74.90±0.21	89.39±0.12
CTX^［7］	Conv-4	72.61±0.21	86.23±0.14
BSNet^［10］	Conv-4	62.84±0.95	85.39±0.56
BDFAN^［12］	Conv-4	79.08±0.20	92.22±0.10
DPGN^［21］	Conv-4	76.05±0.51	89.08±0.38
ProtoNet^［6］	ResNet-12	81.02±0.20	91.93±0.11
FAN^［11］	ResNet-12	84.30±0.18	93.34±0.10
CTX^［7］	ResNet-12	80.39±0.20	91.01±0.11
BDFAN^［12］	ResNet-12	85.44±0.18	94.73±0.09
DPGN^［21］	ResNet-12	75.71±0.47	91.48±0.33
STransGNN		91.08±0.44	94.63±0.50

Table 3. Classification accuracy on Standford Dogs dataset unit: %

View table

Table 3. Classification accuracy on Standford Dogs dataset unit: %

Method	Backbone	5-way 1-shot	5-way 5-shot
ProtoNet^［6］	Conv-4	46.66±0.20	70.77±0.16
RelationNet^［8］	Conv-4	47.35±0.88	66.20±0.74
BSNet^［10］	Conv-4	43.42±0.86	71.90±0.68
FAN^［11］	Conv-4	60.14±0.21	79.26±0.15
CTX^［7］	Conv-4	57.86±0.21	73.59±0.16
BDFAN^［12］	Conv-4	64.74±0.22	81.29±0.14
ProtoNet^［6］	ResNet-12	73.81±0.21	87.39±0.12
FAN^［11］	ResNet-12	76.76±0.21	88.74±0.12
CTX^［7］	ResNet-12	73.22±0.22	85.90±0.13
BDFAN^［12］	ResNet-12	76.89±0.21	88.27±0.12
STransGNN		85.21±0.45	95.68±0.46

Table 4. Classification accuracy on Standford Cars dataset unit: %

View table

Table 4. Classification accuracy on Standford Cars dataset unit: %

Method	Backbone	5-way 1-shot	5-way 5-shot
ProtoNet^［6］	Conv-4	50.88±0.23	74.89±0.18
RelationNet^［8］	Conv-4	46.20±0.91	68.52±0.78
FAN^［11］	Conv-4	67.48±0.22	87.97±0.11
CTX^［7］	Conv-4	66.35±0.21	82.25±0.14
BSNet^［10］	Conv-4	40.89±0.77	86.88±0.50
BDFAN^［12］	Conv-4	75.74±0.20	91.58±0.09
ProtoNet^［6］	ResNet-12	85.46±0.19	95.08±0.08
FAN^［11］	ResNet-12	88.01±0.17	95.75±0.07
CTX^［7］	ResNet-12	85.03±0.19	92.63±0.11
BDFAN^［12］	ResNet-12	90.44±0.15	97.49±0.05
STransGNN		91.10±0.43	94.15±0.47

Table 5. Classification accuracy on Mini-Imagenet dataset unit: %

View table

Table 5. Classification accuracy on Mini-Imagenet dataset unit: %

Method	Backbone	5-way 1-shot	5-way 5-shot
ProtoNet^［6］	ConvNet	49.42±0.78	68.20±0.66
RelationNet^［8］	ConvNet	50.44±0.82	65.32±0.70
GNN^［18］	ConvNet	50.33±0.36	66.41±0.63
DPGN^［21］	ConvNet	66.01±0.36	82.83±0.41
FEAT^［28］	ResNet-12	62.96±0.02	78.49±0.02
MetaGAN^［29］	ResNet-12	52.71±0.64	68.63±0.67
DPGN^［21］	ResNet-12	67.77±0.32	84.60±0.43
STransGNN		71.94±0.53	84.21±0.80

Table 6. Classification accuracy of ablation experiment

View table

Table 6. Classification accuracy of ablation experiment

Experiment No.	Extractor		Measurement method		5-way 1-shot	5-way 5-shot
Experiment No.	ResNet-12	Swin Transformer	Cosine similarity	L2 norm	5-way 1-shot	5-way 5-shot
1	√		√		75.71±0.51	91.48±0.33
2		√	√		88.63±0.45	93.01±0.38
3		√		√	88.20±0.46	92.89±0.60
4		√	√	√	91.08±0.44	94.63±0.50

Table 7. Complexity analysis of models
View table
Table 7. Complexity analysis of models
Method Params /10⁶ FLOPs /10⁹ Memory /GB
GNN^［18］ 1.62 0.19 1.3
DN4^［9］ 0.12 8.98 3.4
DPGN^［21］ 5.79 1.41 1.8
STransGNN 28.66 4.42 2.4

Tools

Get Citation

Copy Citation Text

Kai Wang, Jie Ren, Weichuan Zhang. Few-Shot Image Classification Algorithm of Graph Neural Network Based on Swin Transformer[J]. Laser & Optoelectronics Progress, 2024, 61(12): 1237003

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites