[FIXED] How to realise the 2-D Gaussian filter (like the scipy.ndimage.gaussian_filter) in TensorFlow?

Issue

I am just trying to make a Gaussian filter (like the 'scipy.ndimage.gaussian_filter') to process a 4-D tensor in TensorFlow, the 4-D tensor has a shape of: [16,96,96,3] (16 is the batch size, 96 is the image block size, and 3 is the number of channels). How could I achieve this?

Solution

You just need to create a Gaussian 2D kernel and use a 2D convolution:

import tensorflow as tf

# Make Gaussian kernel following SciPy logic
def make_gaussian_2d_kernel(sigma, truncate=4.0, dtype=tf.float32):
    radius = tf.to_int32(sigma * truncate)
    x = tf.cast(tf.range(-radius, radius + 1), dtype=dtype)
    k = tf.exp(-0.5 * tf.square(x / sigma))
    k = k / tf.reduce_sum(k)
    return tf.expand_dims(k, 1) * k

# Input data
image = tf.placeholder(tf.float32, [16, 96, 96, 3])
# Convolution kernel
kernel = make_gaussian_2d_kernel(5)
# Apply kernel to each channel (see https://stackoverflow.com/q/55687616/1782792)
kernel = tf.tile(kernel[:, :, tf.newaxis, tf.newaxis], [1, 1, 3, 1])
image_filtered = tf.nn.separable_conv2d(
    image, kernel, tf.eye(3, batch_shape=[1, 1]),
    strides=[1, 1, 1, 1], padding='SAME')

Answered By - jdehesa