Opto-Electronic Engineering, Volume. 50, Issue 4, 220232(2023)
Few-shot image classification via multi-scale attention and domain adaptation
Figures & Tables(16)
Fig. 2. The framework of few-shot image classification via multi-scale attention and domain adaptation
Fig. 5. Operating principle of the domain adaptation module. (a) Model training process; (b) Model testing process
Fig. 6. Model convergence on the miniImageNet dataset. (a) 5-way 1-shot；(b) 5-way 5-shot
Fig. 7. Model loss on the miniImageNet dataset. (a) 5-way 1-shot；(b) 5-way 5-shot
Fig. 8. Visualization of features based on convolutional neural network
Fig. 9. Visualization of image features in five classes of miniImageNet. (a) Baseline method; (b) MADA method
Fig. 11. Domain adaptation module analysis. (a) 5-way 1-shot; (b) 5-way 5-shot
Table 1. Structure of the backbone
View tableView in ArticleTable 1. Structure of the backbone
模型结构 输出尺寸 ResNet-12 Conv4 卷积层1 42 42 [3×3，64] × 3 [3×3，64] 卷积层2 21 21 [3×3，160] × 3 [3×3，64] 卷积层3 10 10 [3×3，320] × 3 [3×3，64] 卷积层4 5 5 [3×3，640] × 3 [3×3，64] 池化层 1 1 5×5 Pool 5×5 Pool 参数量 50 MB 0.46 MB
Table 2. Few-shot classification accuracies with 95 confidence interval on the miniImageNet dataset (the number of episodes is 10000)
View tableView in ArticleTable 2. Few-shot classification accuracies with 95 confidence interval on the miniImageNet dataset (the number of episodes is 10000)
模型 骨架网络 5-way 1-shot 5-way 5-shot Matching Net [ 12 ]Conv4 43.56±0.84 55.31±0.37 Proto Net [ 13 ]Conv4 49.42±0.78 68.20±0.66 Relation Net [ 21 ]Conv4 50.44±0.82 65.32±0.70 MAML [ 9 ]Conv4 48.70±1.84 63.11±0.92 DN4 [ 14 ]Conv4 51.24±0.74 71.02±0.64 DSN [ 23 ]Conv4 51.78±0.96 68.99±0.69 BOIL [ 25 ]Conv4 49.61±0.16 66.45±0.37 MADA(ours) Conv4 55.27±0.20 72.12±0.16 Matching Net [ 12 ]ResNet-12 65.64±0.20 78.73±0.15 Proto Net [ 13 ]ResNet-12 60.37±0.83 78.02±0.75 DN4 [ 14 ]ResNet-12 54.37±0.36 74.44±0.29 DSN [ 23 ]ResNet-12 62.64±0.66 78.73±0.45 SNAIL [ 22 ]ResNet-12 55.71±0.99 68.88±0.92 CTM [ 24 ]ResNet-12 64.12±0.82 80.51±0.14 MADA(ours) ResNet-12 67.45±0.20 82.77±0.13
Table 3. Few-shot classification accuracies with 95 confidence interval on the tieredImageNet dataset (the number of episodes is 10000)
View tableView in ArticleTable 3. Few-shot classification accuracies with 95 confidence interval on the tieredImageNet dataset (the number of episodes is 10000)
模型 骨架网络 5-way 1-shot 5-way 5-shot Matching Net [ 12 ]ResNet-12 68.50±0.92 80.60±0.71 Proto Net [ 13 ]ResNet-12 65.65±0.92 83.40±0.65 MetaOpt Net [ 26 ]ResNet-12 65.99±0.72 81.56±0.53 TPN [ 27 ]ResNet-12 59.91±0.94 73.30±0.75 CTM [ 24 ]ResNet-12 68.41±0.39 84.28±1.74 LEO [ 10 ]ResNet-12 66.63±0.05 81.44±0.09 MADA(ours) ResNet-12 70.67±0.22 85.10±0.15
Table 4. Few-shot classification accuracies with 95 confidence interval on the CUB dataset (the number of episodes is 10000)
View tableView in ArticleTable 4. Few-shot classification accuracies with 95 confidence interval on the CUB dataset (the number of episodes is 10000)
模型 骨架网络 5-way 1-shot 5-way 5-shot Matching Net [ 12 ]Conv4 60.52±0.88 75.29±0.75 Proto Net [ 13 ]Conv4 50.46±0.88 76.39±0.64 Relation Net [ 21 ]Conv4 61.10±0.79 76.11±0.69 MAML [ 9 ]Conv4 54.73±0.97 75.75±0.76 Baseline++ [ 28 ]Conv4 60.53±0.83 79.34±0.61 DN4 [ 14 ]Conv4 66.63±0.05 81.44±0.09 MADA(ours) Conv4 62.12±0.24 77.63±0.17
Table 5. Ablation study of few-shot classification accuracies with 95 confidence interval on the miniImageNet (the number of episodes is 10000)
View tableView in ArticleTable 5. Ablation study of few-shot classification accuracies with 95 confidence interval on the miniImageNet (the number of episodes is 10000)
网络 MA DA DE 5-way 1-shot 5-way 5-shot Baseline × × × 60.37±0.83 78.02±0.75 MA √ × × 65.84±0.23 81.94±0.34 MADA √ √ × 67.21±0.18 82.41±0.48 MADA+ √ √ √ 67.45±0.20 82.77±0.13
Tools
Get Citation
Copy Citation Text
Long Chen, Jianlin Zhang, Hao Peng, Meihui Li, Zhiyong Xu, Yuxing Wei. Few-shot image classification via multi-scale attention and domain adaptation[J]. Opto-Electronic Engineering, 2023, 50(4): 220232
Paper Information
Category: Article
Received: Sep. 22, 2022
Accepted: Dec. 29, 2022
Published Online: Jun. 15, 2023
The Author Email: Jianlin Zhang (jlin_zh@163.com)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20