How Can I Efficiently Map Each Pixel Of A Three Channel Image To One Channel?

I am writing a python program to preprocess images to be used as labels for a semantic segmentation task. The original images have three channels, where the vector of three values

Solution 1:

Approach #1

Here's one approach with views and np.searchsorted -

# 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()

# Trace back a 2D array back to given labels
def labelrows(a2D, label_list):
    # Reduce array and labels to 1D
    a1D,b1D = view1D(a2D, label_list)

    # Use searchsorted to trace back label indices
    sidx = b1D.argsort()
    return sidx[np.searchsorted(b1D, a1D, sorter=sidx)]

Hence, to use it for a 3D image array, we need to reshape merging the height and width into one dimension and keeping the color channel dim as it is and use the labeling function.

Approach #2

Tuned for image elements that have [0,255] range, we could leverage matrix-multiplication for the dimensionality-reduction and hence boost up the performance further, like so -

def labelpixels(img3D, label_list):
    # scale array
    s = 256**np.arange(img.shape[-1])

    # Reduce image and labels to 1D
    img1D = img.reshape(-1,img.shape[-1]).dot(s)
    label1D = np.dot(label_list, s)

    # Use searchsorted to trace back label indices
    sidx = label1D.argsort()
    return sidx[np.searchsorted(label1D, img1D, sorter=sidx)]

Sample run on how to extend for image case and also verify -

In [194]: label_list = [[0,255,255], [0,0,0], [0,0,255], [255, 0, 255]]

In [195]: idx = [2,0,3,1,0,3,1,2] # We need to retrieve this back

In [196]: img = np.asarray(label_list)[idx].reshape(2,4,3)

In [197]: img
Out[197]: 
array([[[  0,   0, 255],
        [  0, 255, 255],
        [255,   0, 255],
        [  0,   0,   0]],

       [[  0, 255, 255],
        [255,   0, 255],
        [  0,   0,   0],
        [  0,   0, 255]]])

In [198]: labelrows(img.reshape(-1,img.shape[-1]), label_list)
Out[198]: array([2, 0, 3, 1, 0, 3, 1, 2])

In [217]: labelpixels(img, label_list)
Out[217]: array([2, 0, 3, 1, 0, 3, 1, 2])

Finally, the output would need a reshape back to 2D -

In [222]: labelpixels(img, label_list).reshape(img.shape[:-1])
Out[222]: 
array([[2, 0, 3, 1],
       [0, 3, 1, 2]])