Fine-Grained Lock Cylinder Hole Recognition Based on the Progressive Fusion of Cross-Granularity Features

Fig. 8. Examples of selected images from the four datasets. (a) Images of the CUB-200-2011 dataset; (b) images of the Standford Cars dataset; (c) images of the FGVC-Aircraft dataset; (d) images of the Lock-Hole dataset

Download full size

Fig. 9. Activation maps of convolutional layer classes in the last three stages of the model. (a) Class activation maps of baseline model ; (b) class activation maps of the model which introduces dropout; (c) class activation maps of the model which introduces RSSM

Download full size

Fig. 10. Experimental results of proposed method on CUB-200-2011 dataset. (a) Train and test loss; (b) train and test accuracy

Download full size

Fig. 11. Experimental results of proposed method on the Standford Cars dataset. (a) Train and test loss; (b) train and test accuracy

Download full size

Fig. 12. Experimental results of proposed method on FGVC-Aircraft dataset. (a) Train and test loss curves; (b) train and test accuracy curves

Download full size

Fig. 13. Experimental results of different methods on Lock-Hole dataset. (a) Train loss curves of different methods; (b) test accuracy curves of different methods

Download full size

Table 1. Experimental results of synergistic effects of progressive training, RRCM and progressive training strategies with different stage

View table

Table 1. Experimental results of synergistic effects of progressive training, RRCM and progressive training strategies with different stage

Backbone / training stage / randomized region confusion parameter，p / number of training steps，S	Accuracy	Combined accuracy
ConvNeXt /—/（1）/1	91.2
ConvNeXt /stage 3 /（1，1）/ 2	91.5	91.8
ConvNeXt / stage 3+stage 2 /（1，1，1）/ 3	91.9	92.3
ConvNeXt / stage 3+stage 1 /（1，1，1）/ 3	90.9	91.2
ConvNeXt / stage 3+stage 0 /（1，1，1）/ 3	90.8	91.0
ConvNeXt / stage 2+stage 1 /（1，1，1）/ 3	90.4	90.9
ConvNeXt / stage 1+stage 0 /（1，1，1）/ 3	89.8	90.3
ConvNeXt / stage 3+stage 2+stage 1 /（1，1，1，1）/ 4	91.6	92.1
ConvNeXt / stage 3+stage 2+stage 0 /（1，1，1，1）/ 4	91.5	91.8
ConvNeXt / stage 3+stage 2+stage 1+stage 0 /（1，1，1，1，1）/ 5	91.5	91.7
ConvNeXt / stage 3 /（2，1）/ 2	92.0	92.2
ConvNeXt / stage 3+stage 2 /（4，2，1）/ 3	92.3	92.5
ConvNeXt / stage 3+stage 2 /（8，4，1）/ 3	92.4	92.6
ConvNeXt / stage 3+stage 2 /（8，2，1）/ 3	92.4	92.5
ConvNeXt / stage 2+stage 1 /（8，4，1）/ 3	91.1	91.3
ConvNeXt / stage 1+stage 0 /（8，4，1）/ 3	90.9	91.2
ConvNeXt / stage 3+stage 2+stage 1 /（8，4，2，1）/ 4	91.3	92.0
ConvNeXt / stage 3+stage 2+stage 1 /（16，8，4，1）/ 4	90.8	91.3
ConvNeXt / stage 3+stage 2+stage 1+stage 0 /（16，8，4，2，1）/ 5	90.6	91.2

Table 2. Experimental results of RSSM introducing different hyperparameters
View table
Table 2. Experimental results of RSSM introducing different hyperparameters
Method Parameter Accuracy Combined accuracy
Baseline 92.40 92.62
+Dropout $ρ = 0.5$ 92.47 92.67
+RSSM $ρ = 0.2$ 92.38 92.59
+RSSM $ρ = 0.4$ 92.49 92.68
+RSSM $ρ = 0.6$ 91.57 91.92
+RSSM $ρ \sim U (0.35,0.55)$ 92.58 92.75

Table 3. Experiments results of $L_{c e + m c}$ at different $λ$ , $μ$ uint:%
View table
Table 3. Experiments results of $L_{c e + m c}$ at different $λ$ , $μ$ uint:%
Loss function $λ$ $μ$ Accuracy Combined accuracy
Cross entropy loss function 92.40 92.62
Textual loss function 1.5 10 92.46 92.63
1.5 20 92.51 92.77
2 10 92.54 92.61
2 20 92.52 92.69

Table 4. Results of ablation experiments of proposed method on the Lock-Hole dataset

View table

Table 4. Results of ablation experiments of proposed method on the Lock-Hole dataset

Method	Accuracy	Combined accuracy
ConvNeXt	94.40
Baseline	95.91（+1.51）	96.63（+2.23）
Baseline+RSSM	96.37（+1.97）	96.76（+2.36）
Baseline+ $L_{c e + m c}$	96.40（+2.00）	96.77（+2.37）
Baseline+RSSM+ $L_{c e + m c}$	96.72（+2.32）	97.30（+2.90）

Table 5. Comparative experimental results of different methods on three datasets

View table

Table 5. Comparative experimental results of different methods on three datasets

Method	Backbone	Label	Accuracy /%			GFLOPs
Method	Backbone	Label	CUB-200-2011	Standford Cars	FGVC-Aircraft	GFLOPs
MA-CNN^［21］	VGG-19	Label	86.5	92.8	89.9	61.5
DFL-CNN^［16］	ResNet50	Label	87.4	93.1	91.7	64.8
DCL^［9］	Resnet50	Label	87.8	94.5	93.0	61.5
ACNet^［22］	ResNet50	Label	88.1	94.6	92.4	63.4
AP-CNN^［10］	ResNet50	Label	88.4	95.4	94.1	78.1
SnapMix^［23］	ResNet101	Label	88.5	94.4	93.7	31.5
PMG^［24］	ResNet50	Label	89.6	95.1	93.4	37.4
CAL^［25］	ResNet101	Label	90.6	95.5	94.2	34.6
ViT	ViT-B_16	Label	91.0	93.5	92.1	11.2
DCAL^［26］	R50-ViT	Label	92.0	95.3	93.3	42.4
SIM-Trans^［27］	ViT-B_16	Label	91.8			61.9
TransFG^［28］	ViT-B_16	Label	91.7	94.8		61.9
Proposed method	ConvNeXt	Label	92.8	95.5	94.0	54.7

Table 6. Comparative experimental results of different methods on Lock-Hole dataset

View table

Table 6. Comparative experimental results of different methods on Lock-Hole dataset

Method	Backbone	Label	Accuracy /%	Single recognition time t /s	GFLOPs
ResNet50	ResNet50	Label	93.3	0.007	16.5
ViT	ViT-B_16	Label	92.5	0.011	11.2
ConvNeXt^［19］	ConvNeXt-T	Label	94.4	0.008	17.9
PMG^［24］	ResNet50	Label	95.8	0.012	37.4
DCL^［9］	Resnet50	Label	96.2	0.012	61.5
SnapMix^［23］	ResNet101	Label	95.6	0.015	31.5
SIM-Trans^［27］	ViT-B_16	Label	96.5	0.143	61.9
TransFG^［28］	ViT-B_16	Label	96.7	0.143	61.9
Proposed method	ConvNeXt	Label	97.3	0.016	54.7

Tools

Get Citation

Copy Citation Text

Kunhua Zhu, Lei Sun, Yipeng Liao, Xin Yan, Feifei Cheng. Fine-Grained Lock Cylinder Hole Recognition Based on the Progressive Fusion of Cross-Granularity Features[J]. Laser & Optoelectronics Progress, 2024, 61(18): 1812008

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Instrumentation, Measurement and Metrology

Received: Jan. 2, 2024

Accepted: Mar. 7, 2024

Published Online: Sep. 14, 2024

The Author Email: Yipeng Liao (fzu_lyp@163.com)

DOI:10.3788/LOP240431

CSTR:32186.14.LOP240431

Topics

laser devices and laser physics

Lasers and Laser Optics

Laser physics

laser manufacturing

Instrumentation, Measurement and Metrology

Table 1. Experimental results of synergistic effects of progressive training, RRCM and progressive training strategies with different stage

Table 1. Experimental results of synergistic effects of progressive training, RRCM and progressive training strategies with different stage

Table 2. Experimental results of RSSM introducing different hyperparameters

Table 2. Experimental results of RSSM introducing different hyperparameters

Table 3. Experiments results of Lce+mc at different λ, μ uint:%

Table 3. Experiments results of Lce+mc at different λ, μ uint:%

Table 4. Results of ablation experiments of proposed method on the Lock-Hole dataset

Table 4. Results of ablation experiments of proposed method on the Lock-Hole dataset

Table 5. Comparative experimental results of different methods on three datasets

Table 5. Comparative experimental results of different methods on three datasets

Table 6. Comparative experimental results of different methods on Lock-Hole dataset

Table 6. Comparative experimental results of different methods on Lock-Hole dataset

Table 3. Experiments results of $L_{c e + m c}$ at different $λ$ , $μ$ uint:%

Table 3. Experiments results of $L_{c e + m c}$ at different $λ$ , $μ$ uint:%