Foundations
Image Kernels
How a small matrix of weights (a kernel) slides over a larger grid (an image) to compute local features via dot products.
Stage 1 of 4: The Input
Vertical Edge image, Vertical Edge kernel. Window at row 0, column 0: 0×-1 + 0×0 + 0×1 + 0×-1 + 0×0 + 0×1 + 0×-1 + 0×0 + 0×1 = 0. 16 scan positions in total.
- Input Pixel
- Kernel Window
- Output Cell
Swap the input. A filter only fires on an image that contains its pattern.
An image is just a grid of values. Switch between the two images and read the pixel values — 36 of them, each one an intensity.
Check your understanding
See how small grids of weights slide across an image to detect edges and patterns.
Image Kernels
Before neural networks learn their own filters, traditional computer vision relies on hand-crafted kernels. These small matrices slide across an image to find patterns, like edges or specific textures.
This lab visualizes that mathematical sliding process—the discrete convolution. Watch the dot product update cell by cell to produce a new feature map.
Switch the image and watch the same kernel go silent: a filter is a template, and the dot product only responds when the underlying pattern is actually there.
Reference
- Kernel (Filter)
- A small matrix of weights used to extract features
- Feature Map
- The output image produced by the convolution
- Dot Product
- Element-wise multiplication followed by a sum
- Gain
- The sum of a kernel’s weights. Unity gain (sum = 1) preserves overall brightness; zero gain (sum = 0) is what makes an edge detector return nothing on a flat region.
- Cross-correlation
- A true convolution flips the kernel first; every framework skips the flip because the weights are learned either way. The name stuck.
Break it on purpose
Point a filter at an image that does not contain its pattern — the Horizontal Edge kernel on the Vertical Edge image, or the Vertical Edge kernel on the Horizontal Edge image — and all 16 feature-map cells read exactly 0. Nothing is broken: the filter is reporting truthfully that the feature it looks for is absent.