Related
I'm trying to merge two RGBA images (with a shape of (h,w,4)), taking into account their alpha channels.
Example :
What I've tried
I tried to do this using opencv for that, but I getting some strange pixels on the output image.
Images Used:
and
import cv2
import numpy as np
import matplotlib.pyplot as plt
image1 = cv2.imread("image1.png", cv2.IMREAD_UNCHANGED)
image2 = cv2.imread("image2.png", cv2.IMREAD_UNCHANGED)
mask1 = image1[:,:,3]
mask2 = image2[:,:,3]
mask2_inv = cv2.bitwise_not(mask2)
mask2_bgra = cv2.cvtColor(mask2, cv2.COLOR_GRAY2BGRA)
mask2_inv_bgra = cv2.cvtColor(mask2_inv, cv2.COLOR_GRAY2BGRA)
# output = image2*mask2_bgra + image1
output = cv2.bitwise_or(cv2.bitwise_and(image2, mask2_bgra), cv2.bitwise_and(image1, mask2_inv_bgra))
output[:,:,3] = cv2.bitwise_or(mask1, mask2)
plt.figure(figsize=(12,12))
plt.imshow(cv2.cvtColor(output, cv2.COLOR_BGRA2RGBA))
plt.axis('off')
Output :
So what I figured out is that I'm getting those weird pixels because I used cv2.bitwise_and function (Which btw works perfectly with binary alpha channels).
I tried using different approaches
Question
Is there an approach to do this (While keeping the output image as an 8bit image).
I was able to obtain the expected result in 2 stages.
# Read both images preserving the alpha channel
hh1 = cv2.imread(r'C:\Users\524316\Desktop\Stack\house.png', cv2.IMREAD_UNCHANGED)
hh2 = cv2.imread(r'C:\Users\524316\Desktop\Stack\memo.png', cv2.IMREAD_UNCHANGED)
# store the alpha channels only
m1 = hh1[:,:,3]
m2 = hh2[:,:,3]
# invert the alpha channel and obtain 3-channel mask of float data type
m1i = cv2.bitwise_not(m1)
alpha1i = cv2.cvtColor(m1i, cv2.COLOR_GRAY2BGRA)/255.0
m2i = cv2.bitwise_not(m2)
alpha2i = cv2.cvtColor(m2i, cv2.COLOR_GRAY2BGRA)/255.0
# Perform blending and limit pixel values to 0-255 (convert to 8-bit)
b1i = cv2.convertScaleAbs(hh2*(1-alpha2i) + hh1*alpha2i)
Note: In the b=above the we are using only the inverse alpha channel of the memo image
But I guess this is not the expected result. So moving on ....
# Finding common ground between both the inverted alpha channels
mul = cv2.multiply(alpha1i,alpha2i)
# converting to 8-bit
mulint = cv2.normalize(mul, dst=None, alpha=0, beta=255,norm_type=cv2.NORM_MINMAX, dtype=cv2.CV_8U)
# again create 3-channel mask of float data type
alpha = cv2.cvtColor(mulint[:,:,2], cv2.COLOR_GRAY2BGRA)/255.0
# perform blending using previous output and multiplied result
final = cv2.convertScaleAbs(b1i*(1-alpha) + mulint*alpha)
Sorry for the weird variable names. I would request you to analyze the result in each line. I hope this is the expected output.
You could use PIL library to achieve this
from PIL import Image
def merge_images(im1, im2):
bg = Image.open(im1).convert("RGBA")
fg = Image.open(im2).convert("RGBA")
x, y = ((bg.width - fg.width) // 2 , (bg.height - fg.height) // 2)
bg.paste(fg, (x, y), fg)
# convert to 8 bits (pallete mode)
return bg.convert("P")
we can test it using the provided images:
result_image = merge_images("image1.png", "image2.png")
result_image.save("image3.png")
Here's the result:
Like the image above suggests, how can I convert the image to the left into an array that represent the darkness of the image between 0 for white and decimals for darker colours closer to 1? as shown in the image usingpython 3`?
Update:
I have tried to work abit more on this. There are good answers below too.
# Load image
filename = tf.constant("one.png")
image_file = tf.read_file(filename)
# Show Image
Image("one.png")
#convert method
def convertRgbToWeight(rgbArray):
arrayWithPixelWeight = []
for i in range(int(rgbArray.size / rgbArray[0].size)):
for j in range(int(rgbArray[0].size / 3)):
lum = 255-((rgbArray[i][j][0]+rgbArray[i][j][1]+rgbArray[i][j][2])/3) # Reversed luminosity
arrayWithPixelWeight.append(lum/255) # Map values from range 0-255 to 0-1
return arrayWithPixelWeight
# Convert image to numbers and print them
image_decoded_png = tf.image.decode_png(image_file,channels=3)
image_as_float32 = tf.cast(image_decoded_png, tf.float32)
numpy.set_printoptions(threshold=numpy.nan)
sess = tf.Session()
squeezedArray = sess.run(image_as_float32)
convertedList = convertRgbToWeight(squeezedArray)
print(convertedList) # This will give me an array of numbers.
I would recommend to read in images with opencv. The biggest advantage of opencv is that it supports multiple image formats and it automatically transforms the image into a numpy array. For example:
import cv2
import numpy as np
img_path = '/YOUR/PATH/IMAGE.png'
img = cv2.imread(img_path, 0) # read image as grayscale. Set second parameter to 1 if rgb is required
Now img is a numpy array with values between 0 - 255. By default 0 equals black and 255 equals white. To change this you can use the opencv built in function bitwise_not:
img_reverted= cv2.bitwise_not(img)
We can now scale the array with:
new_img = img_reverted / 255.0 // now all values are ranging from 0 to 1, where white equlas 0.0 and black equals 1.0
Load the image and then just invert and divide by 255.
Here is the image ('Untitled.png') that I used for this example: https://ufile.io/h8ncw
import numpy as np
import cv2
import matplotlib.pyplot as plt
my_img = cv2.imread('Untitled.png')
inverted_img = (255.0 - my_img)
final = inverted_img / 255.0
# Visualize the result
plt.imshow(final)
plt.show()
print(final.shape)
(661, 667, 3)
Results (final object represented as image):
You can use PIL package to manage images. Here's example how it can be done.
from PIL import Image
image = Image.open('sample.png')
width, height = image.size
pixels = image.load()
# Check if has alpha, to avoid "too many values to unpack" error
has_alpha = len(pixels[0,0]) == 4
# Create empty 2D list
fill = 1
array = [[fill for x in range(width)] for y in range(height)]
for y in range(height):
for x in range(width):
if has_alpha:
r, g, b, a = pixels[x,y]
else:
r, g, b = pixels[x,y]
lum = 255-((r+g+b)/3) # Reversed luminosity
array[y][x] = lum/255 # Map values from range 0-255 to 0-1
I think it works but please note that the only test I did was if values are in desired range:
# Test max and min values
h, l = 0,1
for row in array:
h = max([max(row), h])
l = min([min(row), l])
print(h, l)
You have to load the image from the path and then transform it to a numpy array.
The values of the image will be between 0 and 255. The next step is to standardize the numpy array.
Hope it helps.
Having ordered half a dozen webcams online for a project I notice that the colors on the output are not consistent.
In order to compensate for this I have attempted to take a template image and extract the R,G and B histograms and tried to match the target images's RGB histograms based on this.
This was inspired from the description of the solution for a very similar problem Comparative color calibration
The perfect solution will look like this :
In order to try to solve this I wrote the following script which performed poorly:
EDIT (Thanks to #DanMašek and #api55)
import numpy as np
def show_image(title, image, width = 300):
# resize the image to have a constant width, just to
# make displaying the images take up less screen real
# estate
r = width / float(image.shape[1])
dim = (width, int(image.shape[0] * r))
resized = cv2.resize(image, dim, interpolation = cv2.INTER_AREA)
# show the resized image
cv2.imshow(title, resized)
def hist_match(source, template):
"""
Adjust the pixel values of a grayscale image such that its histogram
matches that of a target image
Arguments:
-----------
source: np.ndarray
Image to transform; the histogram is computed over the flattened
array
template: np.ndarray
Template image; can have different dimensions to source
Returns:
-----------
matched: np.ndarray
The transformed output image
"""
oldshape = source.shape
source = source.ravel()
template = template.ravel()
# get the set of unique pixel values and their corresponding indices and
# counts
s_values, bin_idx, s_counts = np.unique(source, return_inverse=True,
return_counts=True)
t_values, t_counts = np.unique(template, return_counts=True)
# take the cumsum of the counts and normalize by the number of pixels to
# get the empirical cumulative distribution functions for the source and
# template images (maps pixel value --> quantile)
s_quantiles = np.cumsum(s_counts).astype(np.float64)
s_quantiles /= s_quantiles[-1]
t_quantiles = np.cumsum(t_counts).astype(np.float64)
t_quantiles /= t_quantiles[-1]
# interpolate linearly to find the pixel values in the template image
# that correspond most closely to the quantiles in the source image
interp_t_values = np.interp(s_quantiles, t_quantiles, t_values)
return interp_t_values[bin_idx].reshape(oldshape)
from matplotlib import pyplot as plt
from scipy.misc import lena, ascent
import cv2
source = cv2.imread('/media/somadetect/Lexar/color_transfer_data/1/frame10.png')
s_b = source[:,:,0]
s_g = source[:,:,1]
s_r = source[:,:,2]
template = cv2.imread('/media/somadetect/Lexar/color_transfer_data/5/frame6.png')
t_b = source[:,:,0]
t_r = source[:,:,1]
t_g = source[:,:,2]
matched_b = hist_match(s_b, t_b)
matched_g = hist_match(s_g, t_g)
matched_r = hist_match(s_r, t_r)
y,x,c = source.shape
transfer = np.empty((y,x,c), dtype=np.uint8)
transfer[:,:,0] = matched_r
transfer[:,:,1] = matched_g
transfer[:,:,2] = matched_b
show_image("Template", template)
show_image("Target", source)
show_image("Transfer", transfer)
cv2.waitKey(0)
Template image :
Target Image:
The Matched Image:
Then I found Adrian's (pyimagesearch) attempt to solve a very similar problem in the following link
Fast Color Transfer
The results seem to be fairly good with some saturation defects. I would welcome any suggestions or pointers on how to address this issue so all web cam outputs could be calibrated to output similar colors based on one template image.
Your script performs poorly because you are using the wrong index.
OpenCV images are BGR, so this was correct in your code:
source = cv2.imread('/media/somadetect/Lexar/color_transfer_data/1/frame10.png')
s_b = source[:,:,0]
s_g = source[:,:,1]
s_r = source[:,:,2]
template = cv2.imread('/media/somadetect/Lexar/color_transfer_data/5/frame6.png')
t_b = source[:,:,0]
t_r = source[:,:,1]
t_g = source[:,:,2]
but this is wrong
transfer[:,:,0] = matched_r
transfer[:,:,1] = matched_g
transfer[:,:,2] = matched_b
since here you are using RGB and not BGR, so the color changes and your OpenCV still thinks it is BGR. That is why it looks weird.
It should be:
transfer[:,:,0] = matched_b
transfer[:,:,1] = matched_g
transfer[:,:,2] = matched_r
As other possible solutions, you may try to look which parameters can be set in your camera. Sometimes they have some auto parameters which you can set manually for all of them to match. Also, beware of this auto parameters, usually white balance and focus and others are set auto and they may change quite a lot in the same camera from one time to another (depending on illumination, etc etc).
UPDATE:
As DanMašek points out, also
t_b = source[:,:,0]
t_r = source[:,:,1]
t_g = source[:,:,2]
is wrong, since the r should be index 2 and g index 1
t_b = source[:,:,0]
t_g = source[:,:,1]
t_r = source[:,:,2]
I have attempted a white patch based calibration routine. Here is the link https://theiszm.wordpress.com/tag/white-balance/.
The code snippet follows:
import cv2
import math
import numpy as np
import sys
from matplotlib import pyplot as plt
def hist_match(source, template):
"""
Adjust the pixel values of a grayscale image such that its histogram
matches that of a target image
Arguments:
-----------
source: np.ndarray
Image to transform; the histogram is computed over the flattened
array
template: np.ndarray
Template image; can have different dimensions to source
Returns:
-----------
matched: np.ndarray
The transformed output image
"""
oldshape = source.shape
source = source.ravel()
template = template.ravel()
# get the set of unique pixel values and their corresponding indices and
# counts
s_values, bin_idx, s_counts = np.unique(source, return_inverse=True,
return_counts=True)
t_values, t_counts = np.unique(template, return_counts=True)
# take the cumsum of the counts and normalize by the number of pixels to
# get the empirical cumulative distribution functions for the source and
# template images (maps pixel value --> quantile)
s_quantiles = np.cumsum(s_counts).astype(np.float64)
s_quantiles /= s_quantiles[-1]
t_quantiles = np.cumsum(t_counts).astype(np.float64)
t_quantiles /= t_quantiles[-1]
# interpolate linearly to find the pixel values in the template image
# that correspond most closely to the quantiles in the source image
interp_t_values = np.interp(s_quantiles, t_quantiles, t_values)
return interp_t_values[bin_idx].reshape(oldshape)
# Read original image
im_o = cv2.imread('/media/Lexar/color_transfer_data/5/frame10.png')
im = im_o
cv2.imshow('Org',im)
cv2.waitKey()
B = im[:,:, 0]
G = im[:,:, 1]
R = im[:,:, 2]
R= np.array(R).astype('float')
G= np.array(G).astype('float')
B= np.array(B).astype('float')
# Extract pixels that correspond to pure white R = 255,G = 255,B = 255
B_white = R[168, 351]
G_white = G[168, 351]
R_white = B[168, 351]
print B_white
print G_white
print R_white
# Compensate for the bias using normalization statistics
R_balanced = R / R_white
G_balanced = G / G_white
B_balanced = B / B_white
R_balanced[np.where(R_balanced > 1)] = 1
G_balanced[np.where(G_balanced > 1)] = 1
B_balanced[np.where(B_balanced > 1)] = 1
B_balanced=B_balanced * 255
G_balanced=G_balanced * 255
R_balanced=R_balanced * 255
B_balanced= np.array(B_balanced).astype('uint8')
G_balanced= np.array(G_balanced).astype('uint8')
R_balanced= np.array(R_balanced).astype('uint8')
im[:,:, 0] = (B_balanced)
im[:,:, 1] = (G_balanced)
im[:,:, 2] = (R_balanced)
# Notice saturation artifacts
cv2.imshow('frame',im)
cv2.waitKey()
# Extract the Y plane in original image and match it to the transformed image
im_o = cv2.cvtColor(im_o, cv2.COLOR_BGR2YCR_CB)
im_o_Y = im_o[:,:,0]
im = cv2.cvtColor(im, cv2.COLOR_BGR2YCR_CB)
im_Y = im[:,:,0]
matched_y = hist_match(im_o_Y, im_Y)
matched_y= np.array(matched_y).astype('uint8')
im[:,:,0] = matched_y
im_final = cv2.cvtColor(im, cv2.COLOR_YCR_CB2BGR)
cv2.imshow('frame',im_final)
cv2.waitKey()
The input image is:
The result of the script is:
Thank you all for suggestions and pointers!!
I'm working on a little problem in my sparetime involving analysis of some images obtained through a microscope. It is a wafer with some stuff here and there, and ultimately I want to make a program to detect when certain materials show up.
Anyways, first step is to normalize the intensity across the image, since the lens does not give uniform lightning. Currently I use an image, with no stuff on, only the substrate, as a background, or reference, image. I find the maximum of the three (intensity) values for RGB.
from PIL import Image
from PIL import ImageDraw
rmax = 0;gmax = 0;bmax = 0;rmin = 300;gmin = 300;bmin = 300
im_old = Image.open("test_image.png")
im_back = Image.open("background.png")
maxx = im_old.size[0] #Import the size of the image
maxy = im_old.size[1]
im_new = Image.new("RGB", (maxx,maxy))
pixback = im_back.load()
for x in range(maxx):
for y in range(maxy):
if pixback[x,y][0] > rmax:
rmax = pixback[x,y][0]
if pixback[x,y][1] > gmax:
gmax = pixback[x,y][1]
if pixback[x,y][2] > bmax:
bmax = pixback[x,y][2]
pixnew = im_new.load()
pixold = im_old.load()
for x in range(maxx):
for y in range(maxy):
r = float(pixold[x,y][0]) / ( float(pixback[x,y][0])*rmax )
g = float(pixold[x,y][1]) / ( float(pixback[x,y][1])*gmax )
b = float(pixold[x,y][2]) / ( float(pixback[x,y][2])*bmax )
pixnew[x,y] = (r,g,b)
The first part of the code determines the maximum intensity of the RED, GREEN and BLUE channels, pixel by pixel, of the background image, but needs only be done once.
The second part takes the "real" image (with stuff on it), and normalizes the RED, GREEN and BLUE channels, pixel by pixel, according to the background. This takes some time, 5-10 seconds for an 1280x960 image, which is way too slow if I need to do this to several images.
What can I do to improve the speed? I thought of moving all the images to numpy arrays, but I can't seem to find a fast way to do that for RGB images.
I'd rather not move away from python, since my C++ is quite low-level, and getting a working FORTRAN code would probably take longer than I could ever save in terms of speed :P
import numpy as np
from PIL import Image
def normalize(arr):
"""
Linear normalization
http://en.wikipedia.org/wiki/Normalization_%28image_processing%29
"""
arr = arr.astype('float')
# Do not touch the alpha channel
for i in range(3):
minval = arr[...,i].min()
maxval = arr[...,i].max()
if minval != maxval:
arr[...,i] -= minval
arr[...,i] *= (255.0/(maxval-minval))
return arr
def demo_normalize():
img = Image.open(FILENAME).convert('RGBA')
arr = np.array(img)
new_img = Image.fromarray(normalize(arr).astype('uint8'),'RGBA')
new_img.save('/tmp/normalized.png')
See http://docs.scipy.org/doc/scipy/reference/generated/scipy.misc.fromimage.html#scipy.misc.fromimage
You can say
databack = scipy.misc.fromimage(pixback)
rmax = numpy.max(databack[:,:,0])
gmax = numpy.max(databack[:,:,1])
bmax = numpy.max(databack[:,:,2])
which should be much faster than looping over all (r,g,b) triplets of your image.
Then you can do
dataold = scip.misc.fromimage(pixold)
r = dataold[:,:,0] / (pixback[:,:,0] * rmax )
g = dataold[:,:,1] / (pixback[:,:,1] * gmax )
b = dataold[:,:,2] / (pixback[:,:,2] * bmax )
datanew = numpy.array((r,g,b))
imnew = scipy.misc.toimage(datanew)
The code is not tested, but should work somehow with minor modifications.
This is partially from FolksTalk webpage:
from PIL import Image
import numpy as np
# Read image file
in_file = "my_image.png"
# convert('RGB') for PNG file type
image = Image.open(in_file).convert('RGB')
pixels = np.asarray(image)
# Convert from integers to floats
pixels = pixels.astype('float32')
# Normalize to the range 0-1
pixels /= 255.0
I have a transparent png image foo.png and I've opened another image with:
im = Image.open("foo2.png")
Now what I need is to merge foo.png with foo2.png.
(foo.png contains some text and I want to print that text on foo2.png)
from PIL import Image
background = Image.open("test1.png")
foreground = Image.open("test2.png")
background.paste(foreground, (0, 0), foreground)
background.show()
First parameter to .paste() is the image to paste. Second are coordinates, and the secret sauce is the third parameter. It indicates a mask that will be used to paste the image. If you pass a image with transparency, then the alpha channel is used as mask.
Check the docs.
Image.paste does not work as expected when the background image also contains transparency. You need to use real Alpha Compositing.
Pillow 2.0 contains an alpha_composite function that does this.
background = Image.open("test1.png")
foreground = Image.open("test2.png")
Image.alpha_composite(background, foreground).save("test3.png")
EDIT: Both images need to be of the type RGBA. So you need to call convert('RGBA') if they are paletted, etc.. If the background does not have an alpha channel, then you can use the regular paste method (which should be faster).
As olt already pointed out, Image.paste doesn't work properly, when source and destination both contain alpha.
Consider the following scenario:
Two test images, both contain alpha:
layer1 = Image.open("layer1.png")
layer2 = Image.open("layer2.png")
Compositing image using Image.paste like so:
final1 = Image.new("RGBA", layer1.size)
final1.paste(layer1, (0,0), layer1)
final1.paste(layer2, (0,0), layer2)
produces the following image (the alpha part of the overlayed red pixels is completely taken from the 2nd layer. The pixels are not blended correctly):
Compositing image using Image.alpha_composite like so:
final2 = Image.new("RGBA", layer1.size)
final2 = Image.alpha_composite(final2, layer1)
final2 = Image.alpha_composite(final2, layer2)
produces the following (correct) image:
One can also use blending:
im1 = Image.open("im1.png")
im2 = Image.open("im2.png")
blended = Image.blend(im1, im2, alpha=0.5)
blended.save("blended.png")
Had a similar question and had difficulty finding an answer. The following function allows you to paste an image with a transparency parameter over another image at a specific offset.
import Image
def trans_paste(fg_img,bg_img,alpha=1.0,box=(0,0)):
fg_img_trans = Image.new("RGBA",fg_img.size)
fg_img_trans = Image.blend(fg_img_trans,fg_img,alpha)
bg_img.paste(fg_img_trans,box,fg_img_trans)
return bg_img
bg_img = Image.open("bg.png")
fg_img = Image.open("fg.png")
p = trans_paste(fg_img,bg_img,.7,(250,100))
p.show()
def trans_paste(bg_img,fg_img,box=(0,0)):
fg_img_trans = Image.new("RGBA",bg_img.size)
fg_img_trans.paste(fg_img,box,mask=fg_img)
new_img = Image.alpha_composite(bg_img,fg_img_trans)
return new_img
Here is my code to merge 2 images of different sizes, each with transparency and with offset:
from PIL import Image
background = Image.open('image1.png')
foreground = Image.open("image2.png")
x = background.size[0]//2
y = background.size[1]//2
background = Image.alpha_composite(
Image.new("RGBA", background.size),
background.convert('RGBA')
)
background.paste(
foreground,
(x, y),
foreground
)
background.show()
This snippet is a mix of the previous answers, blending elements with offset while handling images with different sizes, each with transparency.
the key code is:
_, _, _, alpha = image_element_copy.split()
image_bg_copy.paste(image_element_copy, box=(x0, y0, x1, y1), mask=alpha)
the full function is:
def paste_image(image_bg, image_element, cx, cy, w, h, rotate=0, h_flip=False):
image_bg_copy = image_bg.copy()
image_element_copy = image_element.copy()
image_element_copy = image_element_copy.resize(size=(w, h))
if h_flip:
image_element_copy = image_element_copy.transpose(Image.FLIP_LEFT_RIGHT)
image_element_copy = image_element_copy.rotate(rotate, expand=True)
_, _, _, alpha = image_element_copy.split()
# image_element_copy's width and height will change after rotation
w = image_element_copy.width
h = image_element_copy.height
x0 = cx - w // 2
y0 = cy - h // 2
x1 = x0 + w
y1 = y0 + h
image_bg_copy.paste(image_element_copy, box=(x0, y0, x1, y1), mask=alpha)
return image_bg_copy
the above function supports:
position(cx, cy)
auto resize image_element to (w, h)
rotate image_element without cropping it
horizontal flip