Mathematical filter weighting each pixel against neighbors — foundation for blur, sharpening, edge detection. Every modern image process runs on convolution.
You're in color grading and realize: every blur you apply, every sharpen you pull — it all runs via convolution filters. The math behind it is simple, but its application in the dailies workflow is essential. A convolution filter takes a pixel, looks at its neighbors, and calculates a new value by multiplying each neighbor pixel with a weight. These weights are in a small matrix — usually 3x3 or 5x5 — and determine what happens.
In practice, this means: you need convolution filters to work at all. A blur kernel (blur matrix) simply adds the values of the neighbors evenly — the larger the kernel, the softer the result. A sharpen matrix subtracts the surroundings from the center pixel, thus enhancing edges. For edge detection — crucial for rotoscoping or motion tracking — you use Sobel or Laplace filters, which specifically boost brightness jumps. You don't need to think about this on set itself, but in the VFX department, for denoising or upscaling: convolution runs in the background everywhere.
The crucial point: convolution filters are separable. This means you can filter horizontally and vertically sequentially instead of in 2D — saving enormous processing time. With 4K material in a fast workflow, it's the difference between real-time playback and waiting. Nvidia and AMD have burned convolution into their graphics card shaders; every modern compositor from Nuke to After Effects uses GPU-accelerated convolution. You feel this when your denoising suddenly finishes in 30 seconds instead of three minutes.
Practically: when you design filters yourself — for a specific look or for error correction, for example — you experiment with kernel values. Small values = finer effects, higher values = increased memory consumption and processing time. And be careful: poorly designed kernels can create artifacts or color shifts. Therefore, it's better to use proven presets and only adjust the radius or intensity — the kernel itself remains stable.