Table 3.
Average precision of CNN models with linear SVMs
| Model | Feature layer | Avg | ||
|---|---|---|---|---|
| fc1 | fc2 | fc3 | ||
| AlexNet | 87.14 | 88.48 | 87.29 | 87.64 |
| AlexNet Cropping | 87.91 | 88.93 | 87.99 | 88.28 |
| VGG-16 | 89.20 | 90.02 | 89.65 | 89.62 |
| VGG-16 Cropping | 89.88 | 91.67 | 90.68 | 90.74 |
| GoogLeNet | 85.62 | 86.20 | – | 85.91 |
| GoogLeNet Cropping | 87.57 | 86.83 | – | 87.20 |
| NBI-Net (model A) | 90.93 | 90.64 | 91.21 | 90.93 |
| NBI-Net Aug (model I) | 92.87 | 92.38 | 92.97 | 92.74 |
Layer “fc1” is first fully connected layer before softmax or layer before softmax1 in GoogLeNet, or named “fc6” in AlexNet
Principal component analysis (PCA) here is optional and should be carefully used for dimensionality reduction
The top results have been styled with bold and italic
The best results of comparative group are styled with italic