Incremental Learning Method for Fine-Grained Bird Recognition Based on Prompt Learning

Table 1. Details of datasets
View table
Table 1. Details of datasets
Dataset Number of data Category Training Set Testing Set
CUB-200-2011 11788 200 5994 5794
CIFAR-100 60000 100 50000 10000
5-datasets 261150 50 212785 48365

Table 2. Hyperparameters settings for different datasets
View table
Table 2. Hyperparameters settings for different datasets
Parameter CIFAR-100 5-datasets CUB-200-2011
Optimizer Adam Adam SGD
Learning rate 0.03 0.03 0.03
Momentum 0.9 0.9 0.9
Batch size 16 16 4
Epoch 5 5 10
Random seed 42 42 1993
$λ$ 0.1 0.5 0.1
Prompt length 10 5 5
Top-k 5 4 5

Table 3. Comparison of classification accuracy of different algorithms on CUB-200-2011 dataset

View table

Table 3. Comparison of classification accuracy of different algorithms on CUB-200-2011 dataset

Method	Top-1 accuracy in each task /%
Method	1	2	3	4	5	6	7	8	9	10	$\bar{A}$
Finetune	10.93	13.44	9.67	12.4	12.71	9.25	9.63	8.36	8.08	6.79	10.12
LwF	85.43	60.35	51.93	47.99	46.55	42.94	40.34	39.4	38.96	33.46	48.73
iCaRL	92.71	56.43	41.93	33.41	31.14	24.86	21.45	17.03	15.04	12.81	34.50
EWC	85.43	53.08	57.89	52.34	55.81	51.77	51.43	42.46	39.8	41.69	53.16
L2P	94.33	82.16	72.92	67.57	63.13	61.86	63.86	59.02	57.47	55.81	67.81
Proposed	91.68	85.10	77.80	73.62	70.62	68.60	64.41	61.94	57.51	55.02	70.63

Table 4. Comparison of classification accuracy of different algorithms on CIFAR-100 dataset
View table
Table 4. Comparison of classification accuracy of different algorithms on CIFAR-100 dataset
Method $A_{L}$ $F$
Finetune 5.84 22.62
LwF 60.69 27.77
iCaRL 61.20 27.61
EWC 47.01 33.27
L2P 83.83 7.63
Proposed 84.48 6.43

Table 5. Comparison of classification accuracy of different algorithms on 5-datasets
View table
Table 5. Comparison of classification accuracy of different algorithms on 5-datasets
Method $A_{L}$ $F$
LwF 50.93 34.94
EWC 47.91 38.01
L2P 81.14 4.64
Proposed 82.02 5.34

Table 6. Influence of different text prompts
View table
Table 6. Influence of different text prompts
Method $\bar{A}$ $A_{L}$ $F$
Without family and order 68.27 53.65 30.24
With family 68.95 52.60 41.40
With order 69.90 54.08 30.90
Proposed（with family and order） 70.63 55.02 32.50

Table 7. Comparison of classification accuracy of different algorithms on CIFAR-100 dataset
View table
Table 7. Comparison of classification accuracy of different algorithms on CIFAR-100 dataset
Backbone $A_{L}$ $F$ Parameters /10⁶ FLOPs /10⁹
ViT-T/16 59.35 14.47 5.49 1.34
ViT-S/16 78.57 7.21 21.60 5.30
ViT-B/16 84.48 6.43 85.68 21.09
ViT-L/16 86.13 5.33 303.10 74.74

Table 8. Influence of prompt pool module
View table
Table 8. Influence of prompt pool module
Method $A_{L}$ $F$
ViT 70.29 10.06
ViT+prompt pool 83.30 6.62
ViT+optimized prompt pool 84.48 6.43

Table 9. Influence of prompt length
View table
Table 9. Influence of prompt length
Prompt length CIFAR-100 CUB-200-2011
Learnable parameter $A_{L}$ $F$ Learnable parameter $\bar{A}$ $A_{L}$ $F$
1 92456 81.23 5.34 169356 69.93 56.59 30.63
5 123176 83.45 6.54 200076 70.63 55.02 20.98
10 161576 84.48 6.43 238476 71.36 57.33 30.55

Table 10. Influence of query features
View table
Table 10. Influence of query features
Query feature $A_{L}$ $F$
$h_{c l s}$ 83.70 7.37
$h_{l}$ 84.48 6.43

Tools

Get Citation

Copy Citation Text

Tong Zhu, Haimiao Zhang, Jun Qiu. Incremental Learning Method for Fine-Grained Bird Recognition Based on Prompt Learning[J]. Laser & Optoelectronics Progress, 2024, 61(24): 2437008

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Digital Image Processing

Received: Apr. 3, 2024

Accepted: May. 21, 2024

Published Online: Dec. 17, 2024

The Author Email:

DOI:10.3788/LOP241031

CSTR:32186.14.LOP241031

Topics

laser devices and laser physics

Lasers and Laser Optics

Laser physics

laser manufacturing

Instrumentation, Measurement and Metrology

Table 1. Details of datasets

Table 1. Details of datasets

Table 2. Hyperparameters settings for different datasets

Table 2. Hyperparameters settings for different datasets

Table 3. Comparison of classification accuracy of different algorithms on CUB-200-2011 dataset

Table 3. Comparison of classification accuracy of different algorithms on CUB-200-2011 dataset

Table 4. Comparison of classification accuracy of different algorithms on CIFAR-100 dataset

Table 4. Comparison of classification accuracy of different algorithms on CIFAR-100 dataset

Table 5. Comparison of classification accuracy of different algorithms on 5-datasets

Table 5. Comparison of classification accuracy of different algorithms on 5-datasets

Table 6. Influence of different text prompts

Table 6. Influence of different text prompts

Table 7. Comparison of classification accuracy of different algorithms on CIFAR-100 dataset

Table 7. Comparison of classification accuracy of different algorithms on CIFAR-100 dataset

Table 8. Influence of prompt pool module

Table 8. Influence of prompt pool module

Table 9. Influence of prompt length

Table 9. Influence of prompt length

Table 10. Influence of query features

Table 10. Influence of query features