Watch convolutional filters slide across images and produce feature maps. Explore edge detection, sharpening, and blur kernels: the same operations that power image recognition models.
Sign in to save progress
These are classic hand-crafted kernels, chosen so you can see the mechanics of convolution. In a real CNN the filter values are learned from data during training; nobody writes them by hand. Early layers usually converge to edge-like detectors resembling these; deeper layers combine them into texture and object-part detectors.
3×3 weight matrix applied at each position
Horizontal lines & edges
Detects horizontal boundaries where pixel brightness changes top-to-bottom.
Input (8×8): Checkerboard
Orange box = current 3×3 patch. Stride = 1, no padding, which is why the 8×8 input shrinks to a 6×6 output.
Output (6×6): Horizontal Edge
Green border = current output cell
Computation at position (0,0)
Patch (÷255)
Kernel
Sum
→ Display (0–255)
Element-wise multiply → sum → the whole output map is then min-max rescaled to 0–255 purely for display.
Run convolution to see feature map
Click each layer to see what it detects. Each layer builds on the previous.
Click a layer above to complete this section
Real CNNs stack 10–100+ of these layers. Early layers detect edges, middle layers detect shapes, and deep layers detect complex features like faces or cars. Each layer's output becomes the next layer's input, building from simple to abstract.