问题
Is there a SSIM or even MS-SSIM implementation for TensorFlow?
SSIM (structural similarity index metric) is a metric to measure image quality or similarity of images. It is inspired by human perception and according to a couple of papers, it is a much better loss-function compared to l1/l2. For example, see Loss Functions for Neural Networks for Image Processing.
Up to now, I could not find an implementation in TensorFlow. And after trying to do it by myself by porting it from C++ or python code (such as Github: VQMT/SSIM), I got stuck on methods like applying Gaussian blur to an image in TensorFlow.
Has someone already tried to implement it by himself?
回答1:
After a deep dive into some other python implemention, I could finally implement a running example in TensorFlow:
import tensorflow as tf
import numpy as np
def _tf_fspecial_gauss(size, sigma):
"""Function to mimic the 'fspecial' gaussian MATLAB function
"""
x_data, y_data = np.mgrid[-size//2 + 1:size//2 + 1, -size//2 + 1:size//2 + 1]
x_data = np.expand_dims(x_data, axis=-1)
x_data = np.expand_dims(x_data, axis=-1)
y_data = np.expand_dims(y_data, axis=-1)
y_data = np.expand_dims(y_data, axis=-1)
x = tf.constant(x_data, dtype=tf.float32)
y = tf.constant(y_data, dtype=tf.float32)
g = tf.exp(-((x**2 + y**2)/(2.0*sigma**2)))
return g / tf.reduce_sum(g)
def tf_ssim(img1, img2, cs_map=False, mean_metric=True, size=11, sigma=1.5):
window = _tf_fspecial_gauss(size, sigma) # window shape [size, size]
K1 = 0.01
K2 = 0.03
L = 1 # depth of image (255 in case the image has a differnt scale)
C1 = (K1*L)**2
C2 = (K2*L)**2
mu1 = tf.nn.conv2d(img1, window, strides=[1,1,1,1], padding='VALID')
mu2 = tf.nn.conv2d(img2, window, strides=[1,1,1,1],padding='VALID')
mu1_sq = mu1*mu1
mu2_sq = mu2*mu2
mu1_mu2 = mu1*mu2
sigma1_sq = tf.nn.conv2d(img1*img1, window, strides=[1,1,1,1],padding='VALID') - mu1_sq
sigma2_sq = tf.nn.conv2d(img2*img2, window, strides=[1,1,1,1],padding='VALID') - mu2_sq
sigma12 = tf.nn.conv2d(img1*img2, window, strides=[1,1,1,1],padding='VALID') - mu1_mu2
if cs_map:
value = (((2*mu1_mu2 + C1)*(2*sigma12 + C2))/((mu1_sq + mu2_sq + C1)*
(sigma1_sq + sigma2_sq + C2)),
(2.0*sigma12 + C2)/(sigma1_sq + sigma2_sq + C2))
else:
value = ((2*mu1_mu2 + C1)*(2*sigma12 + C2))/((mu1_sq + mu2_sq + C1)*
(sigma1_sq + sigma2_sq + C2))
if mean_metric:
value = tf.reduce_mean(value)
return value
def tf_ms_ssim(img1, img2, mean_metric=True, level=5):
weight = tf.constant([0.0448, 0.2856, 0.3001, 0.2363, 0.1333], dtype=tf.float32)
mssim = []
mcs = []
for l in range(level):
ssim_map, cs_map = tf_ssim(img1, img2, cs_map=True, mean_metric=False)
mssim.append(tf.reduce_mean(ssim_map))
mcs.append(tf.reduce_mean(cs_map))
filtered_im1 = tf.nn.avg_pool(img1, [1,2,2,1], [1,2,2,1], padding='SAME')
filtered_im2 = tf.nn.avg_pool(img2, [1,2,2,1], [1,2,2,1], padding='SAME')
img1 = filtered_im1
img2 = filtered_im2
# list to tensor of dim D+1
mssim = tf.pack(mssim, axis=0)
mcs = tf.pack(mcs, axis=0)
value = (tf.reduce_prod(mcs[0:level-1]**weight[0:level-1])*
(mssim[level-1]**weight[level-1]))
if mean_metric:
value = tf.reduce_mean(value)
return value
And here is how to run it:
import numpy as np
import tensorflow as tf
from skimage import data, img_as_float
image = data.camera()
img = img_as_float(image)
rows, cols = img.shape
noise = np.ones_like(img) * 0.2 * (img.max() - img.min())
noise[np.random.random(size=noise.shape) > 0.5] *= -1
img_noise = img + noise
## TF CALC START
BATCH_SIZE = 1
CHANNELS = 1
image1 = tf.placeholder(tf.float32, shape=[rows, cols])
image2 = tf.placeholder(tf.float32, shape=[rows, cols])
def image_to_4d(image):
image = tf.expand_dims(image, 0)
image = tf.expand_dims(image, -1)
return image
image4d_1 = image_to_4d(image1)
image4d_2 = image_to_4d(image2)
ssim_index = tf_ssim(image4d_1, image4d_2)
msssim_index = tf_ms_ssim(image4d_1, image4d_2)
with tf.Session() as sess:
sess.run(tf.initialize_all_variables())
tf_ssim_none = sess.run(ssim_index,
feed_dict={image1: img, image2: img})
tf_ssim_noise = sess.run(ssim_index,
feed_dict={image1: img, image2: img_noise})
tf_msssim_none = sess.run(msssim_index,
feed_dict={image1: img, image2: img})
tf_msssim_noise = sess.run(msssim_index,
feed_dict={image1: img, image2: img_noise})
###TF CALC END
print('tf_ssim_none', tf_ssim_none)
print('tf_ssim_noise', tf_ssim_noise)
print('tf_msssim_none', tf_msssim_none)
print('tf_msssim_noise', tf_msssim_noise)
In case you find some errors, please let me know :)
Edit: This implementation only supports gray scaled images
回答2:
This appears to be what you are looking for:
msssim.py
Usage:
python msssim.py --original_image=original.png --compared_image=distorted.png
回答3:
It's a bit late now, but the newer versions of TensorFlow (currently 1.9, 1.10) has an inbuilt function. Check here: TensorFlow MS-SSIM.
You'd need to run it in a Session.
来源:https://stackoverflow.com/questions/39051451/ssim-ms-ssim-for-tensorflow