Some questions about bilinear_kernel.m

[GuoShi28]

Some thing wrong when I use 3 channel bilinear kernel. Following is the 2 upsampling results and original images.


My init code is followings:

function net = add_USN_noise(net, layername, varname, parname, filter_size, scale)
% --------------------------------------------------------------------
crop_default = 1;
stride_default = scale;
net.addLayer(layername, ...
dagnn.ConvTranspose('upsample', stride_default,'crop', crop_default,'hasBias',false), ...
varname{1}, varname{2}, parname);

f = net.getParamIndex(parname) ;
filters = single(bilinear_kernel(4,3,3));
net.params(f).value = filters ;
net.params(f).learningRate = 1;
net.params(f).weightDecay  = 1;
net.params(f).trainMethod = 'adam';

end

Do you have any idea of this problem? Thank you.