bicubic_interp-tensorflow
A differentiable bicubic interpolation module for TensorFlow
Exmple : test.py
tf.image.resize_bicubic doesn't support its gradients for speed issues.
It works different for boundary conditions with tf.image.resize_bicubic.
(It is intended and will not be fixed.)
input : [3, 3]
[[ 0. 1. 2.]
[ 3. 4. 5.]
[ 6. 7. 8.]]
tf.image.resize_bicubic : [9, 9]
[[ 0. 0.40625 1. 1.59375 2. 2.09375]
[ 1.21875 1.625 2.21875 2.8125 3.21875 3.3125 ]
[ 3. 3.40625 4. 4.59375 5. 5.09375]
[ 4.78125 5.1875 5.78125 6.375 6.78125 6.875 ]
[ 6. 6.40625 7. 7.59375 8. 8.09375]
[ 6.28125 6.6875 7.28125 7.875 8.28125 8.375 ]]
bicubic_interp_2d : [9, 9]
[[ 0. 0.32800001 0.78400004 1.21599996 1.67200005 2. ]
[ 0.98400003 1.31200004 1.76800013 2.20000005 2.65600014 2.98399997]
[ 2.352 2.68000007 3.13600016 3.56799984 4.02400017 4.35200024]
[ 3.648 3.97600007 4.43200016 4.86400032 5.31999969 5.64799976]
[ 5.01599979 5.34400034 5.80000019 6.23200035 6.68799973 7.01599979]
[ 6. 6.32800007 6.78399992 7.21600008 7.67199993 8. ]]
Reference
http://blog.demofox.org/2015/08/15/resizing-images-with-bicubic-interpolation/