Efficient Numpy Slicing for Random Image Cropping

For efficient cropping of random 16x16 patches from a 4D Numpy array representing multiple color images (where the first dimension is the number of images, and the second and third are the equal width and height), a strided-based approach can be utilized.

Utilizing np.lib.stride_tricks.as_strided or scikit-image's view_as_windows

These methods create sliding windows as views into the input array, reducing memory overhead. Scikit-image's view_as_windows simplifies the setup by specifying the window shape as a tuple whose elements correspond to the dimensions of the input array. Axes for sliding are assigned window lengths, and other axes are set to 1.

Code Example

# Import scikit-image for view_as_windows
from skimage.util.shape import view_as_windows

# Get sliding windows
w = view_as_windows(X, (1,16,16,1))[...,0,:,:,0]

# Generate random per-image offsets
x = np.random.randint(0,12,X.shape[0])
y = np.random.randint(0,12,X.shape[0])

# Index and extract specific windows
out = w[np.arange(X.shape[0]),x,y]

# Reformat if necessary
out = out.transpose(0,2,3,1)

This code generates four random (x_offset, y_offset) pairs and extracts 4 random 16x16 patches within the given parameters, with minimum memory overhead.