RA-ProtoNet: Classification Based on Meta-Learning for Few-Shot Remote Sensing Scene

Network layer	Structure	Output
Convolution	7 $\times$ 7，64，S=2	112 $\times$ 112，64
Max_pool	3 $\times$ 3，S=2	56 $\times$ 56，64
Residual_lock2_x	$[\begin{array}{l} 3 \times 3,64 \\ 3 \times 3,64 \end{array}] \times 2$	56 $\times$ 56，64
Residual_block3_x	$[\begin{array}{l} 3 \times 3,128 \\ 3 \times 3,128 \end{array}] \times 2$	28 $\times$ 28，128
Residual_block4_x	$[\begin{array}{l} 3 \times 3,256 \\ 3 \times 3,256 \end{array}] \times 2$	14 $\times$ 14，256

Table 2. Experimental environment
View table
Table 2. Experimental environment
Framework Python cuDNN GPU CUDA cuDNN
PyTorch 3.6.2 7.5 RTX 2080TI 10.1 7.5

Table 3. Classification effect on UC-21/AID-30/NP-45 datasets (5-way)

View table

Table 3. Classification effect on UC-21/AID-30/NP-45 datasets (5-way)

Model	Network	UC-21			AID-30			NP-45
Model	Network	1-shot	5-shot	10-shot	1-shot	5-shot	10-shot	1-shot	5-shot	10-shot
Transfer learning	GoogleNet	23.45	45.22	55.59	20.76	40.67	55.63	20.76	40.67	55.63
	AlexNet	20.19	25.08	30.00	20.12	24.56	29.54	20.08	25.49	55.35
	ResNet50	20.95	29.23	46.71	20.07	29.61	45.96	21.14	29.52	49.62
	ResNet101	20.72	29.29	31.61	20.67	23.66	34.90	20.71	27.18	35.18
Meta learning	MAML	47.53	63.13	64.99	47.93	61.79	69.90	42.29	61.84	68.77
	ProtoNet	52.27	69.86	71.69	55.63	68.56	70.48	40.33	63.82	69.53
	Lifelong	39.47	57.40					51.43	72.9
	RS-MetaNet	57.23	76.08	81.23	56.32	74.48	80.57	52.78	71.49	77.37
	RA-ProtoNet	61.77	81.30	84.07	56.74	83.29	87.30	56.60	81.22	86.56

Table 4. Performance evaluation of different embedded networks
View table
Table 4. Performance evaluation of different embedded networks
Network UC-21 /% AID-30 /% NP-45 /% Memory /MB
VGG-16 68.64 55.16 74.34 68.4
ResNet14 81.30 83.29 81.22 35.1
ResNet18 78.92 81.36 82.23 54.9
ResNet50 76.84 77.62 70.38 138

Table 5. Evaluation of different prototype expression methods
View table
Table 5. Evaluation of different prototype expression methods
Feature representation UC-21 /% AID-30 /% NP-45 /%
5-shot 10-shot 5-shot 10-shot 5-shot 10-shot
Sum 76.57 79.78 78.65 79.93 78.61 80.13
Mean 79.49 83.61 81.14 86.45 80.40 84.79
Attention 81.30 84.07 83.29 87.30 81.22 86.56

Tools

Get Citation

Copy Citation Text

Qi He, Jinyuan Zhang, Dongmei Huang, Yanling Du, Huifang Xu. RA-ProtoNet: Classification Based on Meta-Learning for Few-Shot Remote Sensing Scene[J]. Laser & Optoelectronics Progress, 2023, 60(10): 1028003

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites