Split numpy 3-d array into 2-d array of smaller 3-d arrays

Question

I currently have images as numpy array with 3 channels (RGB). I would like to split this efficiently into a 2-d array of smaller 3-d sub arrays. For example if my image is of shape (100, 100, 3), I want to convert it into 10 x 10 array where the elements are (10, 10, 3) images (sub images), while maintaining the spatial orientation. The image height and width will always be equal.

I also wish to reverse the whole operation.

If this is a tough thing to do, is there a way to convert it into a 4-d array in either row or column order, while the element are still the same?

Is there an efficient way to do this using numpy methods?

Answer 1

You can use strides to split your array:

image = np.arange(30000).reshape(100,100,3)

sub_shape = (10,10,3)

#divide the matrix into sub_matrices of subshape
view_shape = tuple(np.subtract(image.shape, sub_shape) + 1) + sub_shape
strides = image.strides + image.strides
sub_matrices = np.squeeze(np.lib.stride_tricks.as_strided(image,view_shape,strides)[::sub_shape[0],::sub_shape[1],:])

sub_matrices shape:

(10, 10, 10, 10, 3)

sub_matrices[i,j,:,:,:] is the [i,j] -th sub-array.

Alternatively you can reshape your image:

sub_shape = (10,10,3)
sub_matrices = np.swapaxes(image.reshape(image.shape[0]/sub_shape[0],sub_shape[0],image.shape[1]/sub_shape[1],sub_shape[1],image.shape[3]), 1, 2)

And again sub_matrices[i,j,:,:,:] is the [i,j] -th sub-array.