python - 分别在 RGB channel 上进行 Tensorflow 2D 卷积？

Question

我想对 RGB 图像应用高斯模糊。我希望它在每个 channel 上独立运行。下面的代码输出具有 3 个 channel 但都具有相同值的模糊图像，导致灰色图像。

gauss_kernel_2d = gaussian_kernel(2, 0.0, 1.0) # outputs a 5*5 tensor
gauss_kernel = tf.tile(gauss_kernel_2d[:, :, tf.newaxis, tf.newaxis], [1, 1, 3, 3]) # 5*5*3*3
image = tf.nn.conv2d(tf.expand_dims(image, 0), gauss_kernel, strides=[1, 1, 1, 1], padding='SAME') # 1*600*800*3
image = tf.squeeze(image) # 600*800*3
# shape of image needs to be [batch, in_height, in_width, in_channels] 
# shape of filter needs to be [filter_height, filter_width, in_channels, out_channels] 

我正在寻找一个 Tensorflow 函数，它分别对每个 R/G/B channel 应用卷积并输出 RGB 模糊图像。

Answer 1

您可以使用 tf.nn.separable_conv2d这样做:

import tensorflow as tf

# ...
gauss_kernel_2d = gaussian_kernel(2, 0.0, 1.0) # outputs a 5*5 tensor
gauss_kernel = tf.tile(gauss_kernel_2d[:, :, tf.newaxis, tf.newaxis], [1, 1, 3, 1]) # 5*5*3*1
# Pointwise filter that does nothing
pointwise_filter = tf.eye(3, batch_shape=[1, 1])
image = tf.nn.separable_conv2d(tf.expand_dims(image, 0), gauss_kernel, pointwise_filter,
                               strides=[1, 1, 1, 1], padding='SAME')
image = tf.squeeze(image) # 600*800*3