0. Basic information

  • Reference information

S. K. Roy, S. Das, T. Song and B. Chanda, “DARecNet-BS: Unsupervised Dual-Attention Reconstruction Network for Hyperspectral Band Selection,” In the IEEE team and Remote Sensing Letters, doi: 10.1109 / LGRS. 2020.3013235.

  • bibtex
@article{Roy2020DARECNETBS, author = {S. K. {Roy} and S. {Das} and T. {Song} and B. {Chanda}}, journal = {IEEE Geoscience and Remote Sensing Letters}, title = {DARecNet-BS: Unsupervised Dual-Attention Reconstruction Network for Hyperspectral Band Selection}, year = {2020}, volume = {}, Number = {}, pages = {1-5}, doi = {10.1109 / LGRS. 2020.3013235}}Copy the code

1. Basic ideas

Similarly, the band selection problem is regarded as the spectral reconstruction problem. The innovation of this paper is that not only the channel attention module is applied, but also the position attention module is added. That’s why it’s called dual-attention, so it can reconstruct a three-digit hyperspectral image data cube.

Overall flow chart

2. Implementation

2.1 Positional attention Module (PAM)

Firstly, three Feature maps A1,A2,A3A_1, A_2, A_3A1,A2,A3 are generated through the convolutional network for the input Patch. The Attention map between two positions in the Patch is calculated using A1,A2A_1, A_2A1 and A2:


q j i p = exp ⁡ ( A 1 . i p . A 2 . j p ) ∑ i . j = 1 V exp ⁡ ( A 1 . i p . A 2 . j p ) q_{ji}^p = \frac{\exp (A_{1, i}^p, A_{2, j}^p)}{\sum_{i,j=1}^V \exp (A_{1, i}^p, A_{2, j}^p)}

Where V represents the number of pixels in a patch.

Then apply the Attention map to pixel J using the following formula:


E P A M . j = Alpha. p ∑ i = 1 V ( q j i p A 3 . i p ) + X j E_{PAM,j} = \alpha^p \sum_{i=1}^V (q_{ji}^p A_{3,i}^p) + X_j

2.2 Band attention module (CAM)

Use a similar calculation method to PAM, but generate a Attention Map between two bands:


q j i c = exp ⁡ ( A 1 . i c . A 2 . j c ) ∑ i . j = 1 B exp ⁡ ( A 1 . i c . A 2 . j c ) q_{ji}^c = \frac{\exp (A_{1, i}^c, A_{2, j}^c)}{\sum_{i,j=1}^B \exp (A_{1, i}^c, A_{2, j}^c)}

Where B represents the number of bands.

The Attention Map applied to the pixels will also change to:


E C A M . j = Alpha. c ∑ i = 1 B ( q j i c A 3 . i c ) + X j E_{CAM,j} = \alpha^c \sum_{i=1}^B (q_{ji}^c A_{3,i}^c) + X_j

2.3 Loss function

The loss function directly calculates the L1 distance between the reconstructed image and the original image:


L 1 ( Theta. b . Theta. e ) = 1 2 N t r ∑ i = 1 N t r ∣ ∣ x i – x ^ i ∣ ∣ 1 \mathcal{L}_1(\theta_b, \theta_e) = \frac{1}{2N_{tr}}\sum_{i=1}^{N_{tr}}||x_i – \hat{x}_i||_1

2.4 Implementation Details

AutoEncoder structure is used for reconstruction network, 3D convolution is used and PReLU is used for activation function.

Input Patch size as 7×7×B7 \times 7 \times B7×7×B, use Cosing to adjust the learning rate, and use diffGrad method as optimizer.

At the end of the band selection, the entropy of the reconstructed output is calculated


H ( b i ) = – ∑ p ( h ) l o g ( p ( h ) ) . s . t . ∑ h p ( h ) = 1 \mathcal{H}(b_i) = – \sum p(h) log(p(h)), s.t. \sum_h p(h)=1

One h is S * SS \ times SS * S Patch in medium frequency histogram Bin number, p (h) = (n (h)/(S (S)) (h) = p (n (h)/(S \ times S)) (h) = p (n (h)/(S (S)) is a probability of h. Finally, the band with the largest top-K entropy is selected as the final band, so the higher entropy usually indicates that the band contains more information.

3. Code

The authors open source the relevant code: github.com/ucalyptus/D…