EE508 · CNN Architectures

2D Convolution on Real Images

Apply real CNN kernels to a photograph and watch the feature map build pixel by pixel. The sliding kernel makes the sliding-window operation viscerally clear — edge detection produces outlines, blur averages neighbors, sharpening amplifies gradients.

📷

Drop image here or click to upload

Or choose a sample:

1 Input Image

2 Kernel

Parameters

Stride (S)

1 2

Padding (P)

None (P=0) Zero-pad (P=1)

Output Size Formula

⌊(W − F + 2P) / S⌋ + 1

Output Display

Grayscale Absolute value

3 Output Feature Map

Press Animate or Full Image to run convolution

Load an image to begin

Live MAC Computation

Patch

Kernel

Current 3×3 Patch (grayscale)

Updates live as kernel sweeps

How It Works

RGB → Grayscale: Y = 0.299R + 0.587G + 0.114B

Each position: multiply patch × kernel element-wise, sum 9 values → 1 output pixel

3-channel image: each RGB channel convolved independently, then summed

Display: output normalized to [0, 255] for visualization