问题
in order to do pattern matching properly convolutions require normalization https://en.wikipedia.org/wiki/Cross-correlation#Normalized_cross-correlation
unfortunately I can't find a way how to make input normalization for conv2d function.
is it hidden in implementation?
回答1:
If I'm not mis-reading that, it's in the image library in TF 1.x:
tf.image.per_image_standardization
https://www.tensorflow.org/api_guides/python/image
By the way, that particular function is a little annoying in that it only takes a single image as an input (3D), but you usually have a 4D tensor representing [batch, height, width, channels] for images. To apply that function to a batch of images you can do this:
imgs4d = tf.map_fn(tf.image.per_image_standardization, imgs4d_float32)
回答2:
tf.image.per_image_standardization does exactly what you want.
Linearly scales image to have zero mean and unit norm.
This op computes (x - mean) / adjusted_stddev, where mean is the average of all values in image, and adjusted_stddev = max(stddev, 1.0/sqrt(image.NumElements())).
stddev is the standard deviation of all values in image. It is capped away from zero to protect against division by 0 when handling uniform images.
You need to do this normalization in a preprocessing step (similar to the place where you would do resizing). Also take a look at other image-related functions.
回答3:
It turned out, I was looking for tf.local_response_normalization (https://www.tensorflow.org/versions/r0.11/api_docs/python/nn/normalization) for some strange reason it goes after conv2 layers and used not very often in examples
来源：https://stackoverflow.com/questions/43700282/is-there-way-to-add-normalization-to-conv2d-in-tensorflow