Knowledge and Geo-Object Based Graph Convolutional Network for Remote Sensing Semantic Segmentation

. 2021 Jun 2;21(11):3848. doi: 10.3390/s21113848

Algorithm 1. For recognizing the spatial correlation.

Input:

m a s k_{i}

of superpixel block i,

m a s k_{j}

of superpixel block j.

Output: spatial correlation between

m a s k_{i}

and

m a s k_{j}

Begin
// use mask to extract feature of $m a s k_{i}$
$f_{i} \leftarrow f c_{1} (m a s k_{i});$ // $f c_{1}$ is a fully connected layer1
// use mask to extract feature of $m a s k_{j}$
$f_{j} \leftarrow f c_{1} (m a s k_{j});$ // $f c_{1}$ is a fully connected layer1
// obtain union mask of $m a s k_{i}$ and $m a s k_{j}$
$m a s k_{i j} \leftarrow u n i o n (m a s k_{i}, m a s k_{j});$ // union is the operation of directly add by pixels
// use union mask to extract feature of mask pair
$f_{i j} \leftarrow f c_{1} (m a s k_{i j});$ // $f c_{1}$ is a fully connected layer1
// obtain feature embedding of correlation $(i, p_{1}, j)$
$V \leftarrow f c_{2} (c o n c a t (f_{i}, f_{i j}, f_{j}));$ // $f c_{2}$ is a fully connected layer2
$f \leftarrow f c_{3} (c o n c a t (f_{i}, V, f_{j})); / / f c_{3}$ is a fully connected layer3
// obtain feature embedding of correlation $(j, p_{2}, i)$
$V^{'} \leftarrow f c_{2} (c o n c a t (f_{j}, f_{i j}, f_{i})); / / f c_{2}$ is a fully connected layer2
$f^{'} \leftarrow f c_{3} (c o n c a t (f_{j}, V^{'}, f_{i})); / / f c_{3}$ is a fully connected layer3
// recognize spatial correlation between i and j
$y \leftarrow a r g m a x (f c_{4} (f));$ $/ / f c_{4}$ is a fully connected layer4
// recognize spatial correlation between j and i
$y^{'} \leftarrow a r g m a x (f c_{4} (f^{'}));$ $/ / f c_{4}$ is a fully connected layer4
Return $y$ , $y^{'}$ ;
End