Change multi dimensional numpy array values according to condition

Question

I have a RGB image stored as a numpy array. I have a color array and these colors I will search in the image and replace those RGB values with a same scalar value. The rest of the RGB values that do not match shall simply be replaced by 0.

The colors I'm searching could be like the following,

colors = []
colors.append((69, 0, 9, 17))
colors.append((196, 127, 128,1))
colors.append((199, 5, 126, 19))
colors.append((55, 127, 126, 4))
colors.append((0, 127, 29, 2))
colors.append((68, 6, 124, 18))

The 4th values on each color is the value that will replace the corresponding RGB values.

I tried using np.asin but it doesn't search for arrays. It only searches for scalars. Right now I am using for loop but its extremely slow.

for i in range(image.shape[0]):
   for j in range(image.shape[1]):
      match = -1
      for k in range(len(colors)):  
         match = k       
         for l in range(3):
            if image[i,j,l] != colors[k][l]:
               match=-1
               break
         if match >=0 :
            break

      val = [0,0,0]
      if match >= 0:
         val = [colors[match][3],colors[match][3],colors[match][3]]
      for l in range(3):
         image[i,j,l] = val[l]

Any efficient approach will be very appreciated.

@b-fg you are right. Its a pseudo code to express what I am trying. Can have some errors. — Tahlil, Oct 25 '18 at 06:25

klim · Accepted Answer · 2018-10-25T09:09:23.883

@Gabriel M

A great approach. But I think that it should be

for r,g,b, replace in colors:

    colors_match = np.where( np.all([image[:,:,0] == r, image[:,:,1] == g, image[:,:,2] == b], axis=0))
    image[colors_match] = replace
    print(colors_match)

or more simply

for r,g,b, replace in colors:

    colors_match = np.all([image[:,:,0] == r, image[:,:,1] == g, image[:,:,2] == b], axis=0)
    image[colors_match] = replace
    print(colors_match)

edited

To replace values that were not converted, keeping the conversion history in another array could be a choice.

converted = np.zeros((image.shape[0], image.shape[1]), dtype=bool)
for r,g,b, replace in colors:

    colors_match = np.all([image[:,:,0] == r, image[:,:,1] == g, image[:,:,2] == b], axis=0)
    image[colors_match] = replace
    converted[colors_match] = True
image[~converted] = 0

Thanks very much it works. But how to replace the other positions that do not match to 0? — Tahlil, Oct 25 '18 at 08:36

score 2 · Answer 2 · answered Oct 25 '18 at 06:30

2

why not simplify your loop like this?:

for r,g,b, replace in colors:

    colors_match = np.where( np.all([image[:,:,0] == r, image[:,:,1] == g, image[:,:,2] == b], axis=0))
    image[colors_match,:] = replace
    print colors_match

answered Oct 25 '18 at 06:30

Gabriel M

1,486
4
17
25

Divakar · Answer 3 · 2018-10-25T06:56:21.300

For ints, here's one way based on dimensionality-reduction discussed in more detail here -

# Based on https://stackoverflow.com/a/38674038/ @Divakar
def matching_index(X, searched_values, invalid_val=-1):
    dims = np.maximum(X.max(0),searched_values.max(0))+1
    X1D = np.ravel_multi_index(X.T,dims)
    searched_valuesID = np.ravel_multi_index(searched_values.T,dims)
    sidx = X1D.argsort()
    sorted_index = np.searchsorted(X1D,searched_valuesID,sorter=sidx)
    sorted_index[sorted_index==len(X1D)] = len(X1D)-1
    idx = sidx[sorted_index]
    valid = X1D[idx] == searched_valuesID
    idx[~valid] = invalid_val
    return valid, idx

# Convert to array
colors = np.asarray(colors)

# Get matching indices and corresponding valid mask
v, idx = matching_index(colors[:,:3],image.reshape(-1,3))
image2D = np.where(v,colors[:,-1][idx],0).reshape(image.shape[:-1])

# If you need a 3D image output
image3D = np.broadcast_to(image2D[...,None], image2D.shape + (3,))

We can also use views to implement equivalent version of matching_index for generic dtype data -

# https://stackoverflow.com/a/45313353/ @Divakar
def view1D(a, b): # a, b are arrays
    a = np.ascontiguousarray(a)
    b = np.ascontiguousarray(b)
    void_dt = np.dtype((np.void, a.dtype.itemsize * a.shape[1]))
    return a.view(void_dt).ravel(),  b.view(void_dt).ravel()

# Based on https://stackoverflow.com/a/38674038/ @Divakar
def matching_index_view(X, searched_values, invalid_val=-1):
    X1D,searched_valuesID = view1D(X,searched_values)
    sidx = X1D.argsort()
    sorted_index = np.searchsorted(X1D,searched_valuesID,sorter=sidx)
    sorted_index[sorted_index==len(X1D)] = len(X1D)-1
    idx = sidx[sorted_index]
    valid = X1D[idx] == searched_valuesID
    idx[~valid] = invalid_val
    return valid, idx

Change multi dimensional numpy array values according to condition

3 Answers3

edited