Graded convolutional networks: scale invariance without the scale and robustness to unknown blur
Abstract
The Gaussian jet of an image is a graded vector space: a dilation by multiplies derivatives of order by . Local graded convolutions are exactly homogeneous differential operators, and a local feature is dilation-invariant if and only if it factors through the scale-normalized jet or, without the scale, through the retraction onto the weighted sphere of the grading. Under unknown Gaussian blur the retraction changes only by a change of scale, whereas the scale-normalized jet acquires a grading factor that no radial normalization removes. At the zero jet the retraction separates image families that radial normalizations and continuous classifiers cannot. Networks built on the retraction are universal for invariant features. On MNIST such a network keeps 88.4% accuracy under a two-pixel blur, against 66.2% for the same network on scale-normalized derivatives; in Lindeberg's Gaussian derivative network the retraction raises accuracy at that blur from 91.5% to 95.1%.
Figures & tables
| Model | retr. | gate | var. | var. | const. | |||||
| GCN | ||||||||||
| SNT | ||||||||||
| UCN | ||||||||||
| GCN , equal-grades gate | 95.5 | 94.3 | 95.9 | 94.1 | 95.9 | 0.017 | 0.018 | 89 | ||
| UCN , graded gate | 76.1 | 91.6 | 96.5 | 89.6 | 76.3 | 0.090 | 0.042 | 52 | ||
| CNN, size-normalized | – | – | 89.9 | 91.9 | 92.1 | 92.1 | 92.1 | 0.004 | 0.002 | 93 |
| Model | test | zoom | zoom | orbit var. | |
|---|---|---|---|---|---|
| GCN | 95.3 | 95.7 | 95.5 | 0.0110 | |
| UCN | 95.3 | 95.9 | 82.2 | 0.0338 | |
| wrong | 93.8 | 94.1 | 83.1 | 0.0356 | |
| wrong | 95.0 | 95.3 | 92.3 | 0.0246 | |
| learned | 95.6 | 96.2 | 89.5 | 0.0217 | |
| learned, grade l.r. | 95.7 | 96.3 | 89.4 | 0.0205 |
| Model | test | zoom | |||
|---|---|---|---|---|---|
| GCN , , full readout | 95.3 | 95.5 | 71.7 | 24.8 | 12.2 |
| GCN , | 94.3 | 94.3 | 67.3 | 18.3 | 9.3 |
| GCN , no product layer | 92.1 | 92.5 | 67.7 | 22.6 | 11.7 |
| GCN , invariant readout | 88.7 | 87.1 | 86.1 | 86.1 | 87.1 |
| UCN , full readout | 95.3 | 82.2 | 66.6 | 15.7 | 8.5 |
| UCN , invariant readout | 88.1 | 51.1 | 81.6 | 81.7 | 81.9 |
| zoom , blur | res. | zoom , | zoom , | |||||||
| Model | ||||||||||
| GCN | ||||||||||
| SNT | ||||||||||
| UCN | ||||||||||
| CNN, size-norm. | 92.1 | 91.8 | 88.4 | 77.4 | 60.1 | 89.9 | 89.9 | 55.1 | 92.1 | 91.7 |
| more accurate | SNT 3/3 | SNT 2/3 | GCN 2/3 | GCN 3/3 | GCN 3/3 | SNT 1/3 | SNT 2/3 | GCN 3/3 | SNT 3/3 | SNT 2/3 |
| zoom , blur | zoom, | ||||||
| Model | , | , | |||||
| without the second contrast normalization | |||||||
| GCN | 95.8 | 95.8 | 95.4 | 93.6 | 88.5 | 84.9 | 95.8 |
| SNT | 96.8 | 96.7 | 95.4 | 84.4 | 64.7 | 59.5 | 96.7 |
| gates exchanged, contrast normalized as in Tab. 4 | |||||||
| GCN , equal-grades gate | 95.9 | 96.0 | 95.5 | 93.6 | 88.0 | 84.8 | 95.9 |
| Pooling | var. | |||||||
|---|---|---|---|---|---|---|---|---|
| eight scale channels | ||||||||
| maximum over channels [ 19 ] | 69.3 | 96.6 | 98.0 | 91.4 | 84.1 | 81.6 | 86.8 | 0.122 |
| image-frame pooling | 95.9 | 98.1 | 98.5 | 98.5 | 98.0 | 97.9 | 97.5 | 0.003 |
| image-frame pooling, graded first layer | 96.8 | 98.3 | 98.8 | 98.8 | 98.8 | 98.1 | 98.1 | 0.004 |
| image-frame pooling, retraction only | 96.4 | 98.9 | 98.6 | 98.5 | 97.9 | 97.8 | 97.9 | 0.004 |
| four scale channels | ||||||||
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
| Check | Result |
|---|---|
| E1 sampled Gaussian derivatives vs. the analytic jet of a blob, | |
| E2 scale-space covariance ( 2 ) under , analytic images, | |
| E3 position features under for / / | |
| E3 features of / under the dilation action | 0.95 / 1.14 |
| E4 harmonic coordinates transform by ; invariant readout rotation-invariant | ; |
| E4 pooled vector, invariant readout, under exact rotations of digits, GCN / UCN | / |
Convolutional Neural Networks
Convolutional neural networks (CNNs) learn spatial and temporal patterns using shared filters, with applications from image recognition to scientific prediction.
See all 439 papers