Fine-Grained Image Classification Model Based on Improved Transformer

Table 1. Detailed information of fine-grained image datasets
View table
Table 1. Detailed information of fine-grained image datasets
Dataset Number of subclasses Number of samples on training set Number of samples on test set
CUB-200-2011 200 5994 5794
Stanford Cars 120 8144 8041
Stanford Dogs 196 12000 8580

Table 2. Ablation experiment analysis of MEA and FS module
View table
Table 2. Ablation experiment analysis of MEA and FS module
No. Method Model composition Accuracy /%
1） ViT（baseline） SA 85.8
2） ViT（EA） EA 86.9
3） ViT（FS） SA+FS 86.6
4） ViT（EA&FS） EA+FS 87.9

Table 3. Ablation experiment analysis of contrastive loss
View table
Table 3. Ablation experiment analysis of contrastive loss
No. Method Model composition Accuracy /%
1） ViT（baseline） SA 85.8
5） ViT（SA&L_CON） SA+L_CON 86.2
4） ViT（EA&FS） EA+FS 87.9
6） ViT（EA&FS&L_CON） EA+FS+L_CON 88.2

Table 4. Ablation experiment analysis of distillation loss

View table

Table 4. Ablation experiment analysis of distillation loss

Method	Model composition	Teacher	Accuracy /%
ViT（EA&FS&L_CON）	EA+FS+L_CON		88.2
TransFC（EA&FS&L_CON&L_DIS1）	EA+FS+L_CON+L_DIS	VGG-16	88.9
TransFC（EA&FS&L_CON&L_DIS2）	EA+FS+L_CON+L_DIS	ResNet-50	89.1
TransFC（EA&FS&L_CON&L_DIS3）	EA+FS+L_CON+L_DIS	ResNet-101	89.5
TransFC（EA&FS&L_CON&L_DIS4）（ours）	EA+FS+L_CON+L_DIS	DenseNet-121	89.8

Table 5. Experiment comparison of different weakly supervised fined-grained image classification methods

View table

Table 5. Experiment comparison of different weakly supervised fined-grained image classification methods

Method	Base model	Accuracy /%
Method	Base model	CUB-200-2011	Stanford Dogs	Stanford Cars
DB^［24］	ResNet-50	88.6	87.7	94.9
SEF^［25］	ResNet-50	87.3	88.8	94.0
B-CNN^［4］	VGG-16	84.1		91.3
WS-DAN^［26］	Inception V3	89.4	90.0	94.1
ACNet^［27］	VGG-16	87.8		94.3
ACNet^［27］	ResNet-50	88.1		94.6
MAMC^［28］	ResNet-101	86.5	85.2	93.0
DVAN^［6］	VGG-16	79.0	81.5	87.1
RA-CNN^［5］	VGG-19	85.5	87.3	92.5
MC Loss^［29］	ResNet-50	87.3		93.7
ViT^［11］	SA	85.8	87.2	92.6
TransFC（ours）	EA	89.8	90.2	94.7

Tools

Get Citation

Copy Citation Text

Zhansheng Tian, Libo Liu. Fine-Grained Image Classification Model Based on Improved Transformer[J]. Laser & Optoelectronics Progress, 2023, 60(2): 0210013

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites