Fine-Grained Image Classification Based on Feature Fusion and Ensemble Learning

Wenli Zhang; Wei Song

doi:10.3788/LOP240759

Laser & Optoelectronics Progress, Volume. 61, Issue 22, 2237010(2024)

Fine-Grained Image Classification Based on Feature Fusion and Ensemble Learning

Wenli Zhang^1,2 and Wei Song^1,2、*

Author Affiliations

¹School of Artificial Intelligence and Computer Science, Jiangnan University, Wuxi 214122, Jiangsu , China

²Jiangsu Provincial Engineering Laboratory of Pattern Recognition and Computational Intelligence, Jiangnan University, Wuxi 214122, Jiangsu , China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(17)

Fig. 1. Illustration of different mixup methods. (a) Input; (b) Cutout; (c) Mixup; (d) CutMix; (e) AGMix

Download full size

Fig. 2. Framework of ViT

Download full size

Fig. 3. Framework of FFEL-Net

Download full size

Fig. 4. Framework of part relationship modeling module

Download full size

Fig. 5. Framework of multi-expert part voting module

Download full size

Fig. 6. PR curves of different ablation methods. (a) CUB-200-2011; (b) Stanford Dogs; (c) NAbirds; (d) IP102

Download full size

Fig. 7. Ablation results with different γ and η values. (a) CUB-200-2011; (b) Stanford Dogs; (c) NABirds; （d） IP102

Download full size

Fig. 8. Visualization results with different γ values

Download full size

Fig. 9. t-SNE visualization results on four datasets. (a) CUB-200-2011; (b) Stanford Dogs; (c) NABirds; (d) IP102

Download full size

Fig. 10. Feature visualization results. (a) CUB-200-2011; (b) Stanford Dogs; (c) NABirds; (d) IP102

Download full size

Table 1. Introduction of four datasets
View table
Table 1. Introduction of four datasets
Dataset Calss Train Validation
CUB-200-2011 200 5994 5794
Stanford Dogs 120 12000 8000
NABirds 555 23929 24633
IP102 102 45095 22619

Table 2. Ablation results of different modules on four datasets

View table

Table 2. Ablation results of different modules on four datasets

MLFF	MPV	AGMix	CUB-200-2011	Stanford Dogs	NABirds	IP102
-	-	-	90.52	91.57	89.93	73.41
√	-	-	91.25	92.34	90.32	74.89
-	√	-	91.16	92.18	90.25	74.62
-	-	√	91.39	92.42	90.46	74.98
√	√	-	91.53	92.68	90.57	75.57
√	-	√	91.77	92.83	90.82	75.84
-	√	√	91.66	92.76	90.64	75.74
√	√	√	91.92	93.10	90.98	76.21

Table 3. Ablation results with different K values

View table

Table 3. Ablation results with different K values

Hyper-parameter	Accuracy /%				Parameters /10⁶
Hyper-parameter	CUB-200-2011	Standford Dogs	NABirds	IP102	Parameters /10⁶
K=2	91.53	92.70	90.50	75.74	93.21
K=3	91.92	93.10	90.98	76.21	93.79
K=4	91.66	92.83	90.62	75.83	94.38
K=5	91.35	92.52	90.35	75.44	94.97
K=6	91.09	92.30	90.22	75.07	95.57

Table 4. Comparative results of different mixup methods
View table
Table 4. Comparative results of different mixup methods
Method CUB-200-2011 Stanford Dogs NABirds IP102
ViT 90.52 91.57 89.93 73.41
+Cutout 90.74 91.86 90.01 73.97
+ Mixup 91.12 92.15 90.22 74.48
+ CutMix 91.28 92.23 90.31 74.75
+AGMix 91.39 92.42 90.46 74.98

Table 5. Comparison results on CUB-200-2011, Standford Dogs and NABirds datasets

View table

Table 5. Comparison results on CUB-200-2011, Standford Dogs and NABirds datasets

Method	Backbone	CUB-200-2011	Standford Dogs	NABirds
API-Net^［27］	ResNet-50	87.7	88.3	86.2
FDL^［7］	ResNet-50	88.6	85.0
MGE-CNN^［9］	ResNet-50	88.5		88.0
MRDMN^［10］	ResNet-50	88.8	89.1
PMG-V2^［28］	ResNet-50	89.9	89.1	87.6
ViT^［11］	ViT-B/16	90.5	91.6	89.9
TransFG^［12］	ViT-B/16	91.7	92.3	90.8
TPSKG^［14］	ViT-B/16	91.3	92.5	90.3
FFVT^［29］	ViT-B/16	91.6	91.5
SM-ViT^［30］	ViT-B/16	91.6	92.3	90.5
AA-Trans^［15］	ViT-B/16	91.4	90.8	90.2
IELT^［5］	ViT-B/16	91.8	91.8	90.8
FFEL-Net	ViT-B/16	91.92	93.10	90.98

Table 6. Comparison results on IP102 dataset

View table

Table 6. Comparison results on IP102 dataset

Method	Backbone	IP102 /%
API-Net^［27］	ResNet-50	56.9
DMF-ResNet^［31］	ResNet-50	59.2
SGDL-Net^［32］	DenseNet121	72.7
MS-ALN+DL^［33］	ResNet-50	74.6
ViT^［11］	ViT-B/16	73.4
FFVT^［29］	ViT-B/16	74.6
TransFG^［12］	ViT-B/16	74.8
AATrans^［15］	ViT-B/16	75.0
IELT^［5］	ViT-B/16	75.4
FFEL-Net	ViT-B/16	76.21

Table 7. Analysis of intensity values before

View table

Table 7. Analysis of intensity values before

Method	backbone	Parameter /10⁶	Inference time /s	FLOPs /10⁹	ChinaBirds /%
ViT^［11］	ViT-B/16	86.42	0.011	67.46	95.86
TransFG^［12］	ViT-B/16	86.42	0.147	108.40	97.35
FFVT^［29］	ViT-B/16	86.42	0.010	67.73	97.12
AA-Trans^［15］	ViT-B/16	86.42	0.011	65.97	96.53
IELT^［5］	ViT-B/16	99.98	0.013	72.60	97.09
FFEL-Net	ViT-B/16	93.79	0.012	68.74	97.83

Tools

Get Citation

Copy Citation Text

Wenli Zhang, Wei Song. Fine-Grained Image Classification Based on Feature Fusion and Ensemble Learning[J]. Laser & Optoelectronics Progress, 2024, 61(22): 2237010

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites