Pixels & neighborhoods

Build a Filter

Edit a kernel and see what it does to an image. Inspect one neighborhood to understand every multiply and add.

10 min to exploreComputer Vision

01 / UNDERSTAND

How can nine numbers change the appearance of an image?

A filter combines nearby pixel values. Kernel weights determine the operation, border handling affects edge pixels, and display clipping can hide values outside the visible range.

02 / EXPLORE

Make a prediction.
Then move a control.

Nine numbers. A different image.

Build a filter

Edit a 3 × 3 kernel and follow the arithmetic for one pixel.

Kernel preset

Kernel weights

Kernel row 1, column 1Kernel row 1, column 2Kernel row 1, column 3Kernel row 2, column 1Kernel row 2, column 2Kernel row 2, column 3Kernel row 3, column 1Kernel row 3, column 2Kernel row 3, column 3

Divide weighted sum byBeyond the image border

At the edge, repeat the nearest valid pixel. The output stays 32 × 32.

Convention: cross-correlation. The kernel slides without being flipped, as in PyTorch Conv2d. Mathematical convolution flips the kernel first.

Original32 × 32 grayscale pixels

FilteredSame size · clipped to 0–255

Click either image to inspect a pixel. With an image focused, use the arrow keys. The blue box marks the selected output; the outlined neighborhood shows its source pixels.

Column x 13Row y 10

Weighted sum—

Raw output—

Displayed pixel—

Divide, clip to [0, 255], then round for display.

Inside one output pixel

Pixel (13, 10)

Each cell shows source × weight = contribution. The kernel stays in the order you entered.

Sum the nine contributions, then divide.

An educational experiment with illustrative data. The controls run locally in your browser.

03 / SEE THE CODE

The idea,
in Python.

A compact implementation of the core concept. Open the code when you are ready to connect the visual to the algorithm.

Open the core implementation Python ↗

import numpy as np

image = np.array([[0., 0., 0.], [0., 255., 0.], [0., 0., 0.]])
kernel = np.ones((3, 3)) / 9
padded = np.pad(image, 1, mode="constant")
raw = np.zeros_like(image)

for row in range(image.shape[0]):
    for col in range(image.shape[1]):
        patch = padded[row:row + 3, col:col + 3]
        raw[row, col] = np.sum(patch * kernel)

# Unflipped kernel: cross-correlation, as used in many CNN APIs.
display = np.clip(raw, 0, 255).astype(np.uint8)
print(raw)

04 / TRY IT

Explain the change.
Test your reasoning.

Compare blur, edge and sharpen kernels on the same image.
Inspect a pixel near an edge and explain its raw filtered value.
Change the border rule. Why do image boundaries behave differently?

← All visual labs Next: Neural Network Parameter Explorer