Image preprocessing — scale and standardize

Background

Before an image enters a network it's almost always normalized: scale the raw $[0, 255]$ pixels to $[0, 1]$, then standardize each channel with a per-channel mean and std (the ImageNet recipe):

$$\text{out}_{\cdot,\cdot,c} = \frac{\dfrac{\text{img}_{\cdot,\cdot,c}}{255} - \mu_c}{\sigma_c}$$

This centers and rescales the inputs so optimization behaves well.

Problem statement

Implement normalize_image(img, mean, std) that scales by 255 then standardizes per channel.

Input

• img — array-like of shape (H, W, C), raw pixel values in $[0, 255]$.
• mean — array-like of length C, per-channel means (in $[0,1]$ scale).
• std — array-like of length C, per-channel std deviations.

Output

Returns an np.ndarray of shape (H, W, C): the normalized image.

Examples

Example 1

Input:  img = full 2×2×3 of 255, mean = [0.5, 0.5, 0.5], std = [0.5, 0.5, 0.5]
Output: all 1.0

Explanation: $255/255 = 1$; $(1 - 0.5)/0.5 = 1$ for every channel.

Example 2 — black pixels

Input:  img = full 2×2×3 of 0, mean = [0.5, 0.5, 0.5], std = [0.5, 0.5, 0.5]
Output: all -1.0

Explanation: $(0 - 0.5)/0.5 = -1$.

Constraints

• Divide by 255 first, then subtract mean and divide by std, per channel (broadcast over H, W).
• Return shape (H, W, C).

Notes

• mean and std broadcast across the spatial dimensions because they have length C matching the last axis.