Dual-attention random selection global context fine-grained recognition network

Table 1. Dataset parameters table
View table
View in Article
Table 1. Dataset parameters table
数据集 Stanford-cars CUB-200-2011 FGVC-Aircraft VMRURS
训练样本量 8 144 5 994 6 667 3 096
测试样本量 8 041 5 794 3 333 751
类别数 196 200 100 48

Table 2. Variant structure of DRC-Net

View table

View in Article

Table 2. Variant structure of DRC-Net

网络变体模块	相关描述
Baseline	基线网络ConvNeXt
DRSM1	Stage1后嵌入双注意力随机选择模块
DRSM2	Stage2后嵌入双注意力随机选择模块
DRSM3	Stage3后嵌入双注意力随机选择模块
DRSM4	Stage4后嵌入双注意力随机选择模块
GCAM	全局上下文注意力模块

Table 3. Comparison of ablation experiments with DRSM embedded in different locations
View table
View in Article
Table 3. Comparison of ablation experiments with DRSM embedded in different locations
Method Accuracy/%
Baseline 93.8
Baseline+DRSM1 94.2（+0.4）
Baseline+DRSM2 94.6（+0.8）
Baseline+DRSM3 94.6（+0.8）
Baseline+DRSM4 94.3（+0.5）
Baseline+DRSM2+DRSM3 94.9（+1.1）

Table 4. Different modules are added to the network

View table

View in Article

Table 4. Different modules are added to the network

Method	Accuracy/%	Params/M	FLOPs/G
Baseline	93.8	87.5	61.4
Baseline+DRSM	94.9（+1.1）	87.5（+0）	65.8（+4.4）
Baseline+GCAM	94.1（+0.3）	88.7（+1.2）	62.2（+0.8）
Ours（Baseline+DRSM+GCAM）	95.2（+1.4）	88.7（+1.2）	66.6（+5.2）

Table 5. Experiment results of the other methods based on ConvNeXt backbone network and ours
View table
View in Article
Table 5. Experiment results of the other methods based on ConvNeXt backbone network and ours
Method Accuracy/% Params/M FLOPs/G
ConvNeXt 93.8 87.5 61.4
FBSD 94.6 97.5 82.3
PMG 95.1 97.5 82.2
WS-DAN 94.6 97.2 63.8
AC-Net 94.8 89.4 86.3
SeA 95.1 108.2 93.2
Ours 95.2 88.7 66.6

Table 6. Experimental effects of different networks on three open fine-grained datasets

View table

View in Article

Table 6. Experimental effects of different networks on three open fine-grained datasets

Number	Method	Backbone	Accurcy/%			Params/M	FLOPs/G
Number	Method	Backbone	Car	Bird	Air	Params/M	FLOPs/G
1	RA-CNN	VGG	92.5	85.3	87.8	32.1	38.6
2	ResNet50	ResNet50	91.5	86.2	86.8	23.5	33.5
3	WS-DAN	Inceptionv3	94.5	89.4	93.0	33.2	18.5
4	PMG	ResNet50	95.1	89.6	93.4	45.1	37.4
5	Corss-X	ResNet50	94.6	87.7	91.3	25.6	38.9
6	FBSD	Densenet161	94.5	89.8	93.2	46.8	53.1
7	API-Net	ResNet101	94.9	88.6	93.0	46.1	63.0
8	AC-Net	Resnet50	94.6	88.1	92.4	44.2	69.6
9	ViT	ViT-B_16	93.5	90.2	93.1	86.1	62.0
10	TransFG	ViT-B_16	94.8	91.7	93.4	86.4	62.0
11	FFVT	ViT-B_16	94.6	91.6	93.2	86.2	62.0
12	ConvNeXt	ConvNeXt	93.8	91.2	93.2	87.5	61.4
13	Ours	ConvNeXt	95.2	92.1	94.0	88.7	66.6

Table 7. Experimental effects of different networks on VMRURS dataset

View table

View in Article

Table 7. Experimental effects of different networks on VMRURS dataset

方法	主干网络	Accuracy/%	Params/M	FLOPs/G
ResNet50	ResNet50	90.6	23.5	33.5
PMG	ResNet50	95.2	45.1	37.4
FBSD	Densenet161	95.1	46.8	53.1
WS-DAN	Inceptionv3	95.8	33.2	18.5
API-Net	ResNet101	95.9	46.1	63.0
AC-Net	ResNet50	95.8	44.2	69.6
ViT	ViT-B_16	95.1	86.1	62.0
TransFG	ViT-B_16	96.3	86.4	62.0
FFVT	ViT-B_16	96.2	86.2	62.0
ConvNeXt	ConvNeXt	95.5	87.5	61.4
Ours	ConvNeXt	97.0	88.7	66.6

Tools

Get Citation

Copy Citation Text

Shengjun XU, Yang JING, Zhongxing DUAN, Minghai LI, Haitao LI, Fuyou LIU. Dual-attention random selection global context fine-grained recognition network[J]. Chinese Journal of Liquid Crystals and Displays, 2024, 39(4): 506

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Research Articles

Received: Mar. 31, 2023

Accepted: --

Published Online: May. 28, 2024

The Author Email: Yang JING (jingyang0525@xauat.edu.cn)

DOI:10.37188/CJLCD.2023-0114

Topics

Table 1. Dataset parameters table

Table 1. Dataset parameters table

Table 2. Variant structure of DRC-Net

Table 2. Variant structure of DRC-Net

Table 3. Comparison of ablation experiments with DRSM embedded in different locations

Table 3. Comparison of ablation experiments with DRSM embedded in different locations

Table 4. Different modules are added to the network

Table 4. Different modules are added to the network

Table 5. Experiment results of the other methods based on ConvNeXt backbone network and ours

Table 5. Experiment results of the other methods based on ConvNeXt backbone network and ours

Table 6. Experimental effects of different networks on three open fine-grained datasets

Table 6. Experimental effects of different networks on three open fine-grained datasets

Table 7. Experimental effects of different networks on VMRURS dataset

Table 7. Experimental effects of different networks on VMRURS dataset