image features and edges
/ IM-ij FEE-churs and EJ-iz /
Before a machine can recognize a face or a road, it has to find something to grab onto in the raw sea of pixels. Features are those handholds: edges (sharp changes in brightness, like the outline where a dark roof meets bright sky), corners, blobs, and textures. An edge is the most basic feature — it marks a boundary. A corner, where two edges meet, is even more useful because it is easy to relocate in another photo. These small, distinctive landmarks are the alphabet from which higher-level recognition is spelled out.
For decades, researchers hand-designed clever recipes to detect these features — names you may meet include the Canny edge detector, and keypoint methods like SIFT and ORB. The idea was to boil a messy photo down to a compact set of stable landmarks, then match those landmarks across images. This is how panorama-stitching glues your photos together, how a phone tracks the world for augmented reality, and how older recognition systems worked before deep learning.
Here is the honest historical turn. Modern deep networks rarely use these hand-crafted features directly; instead they learn their own features from data, and the early layers of a trained network end up rediscovering edge-and-corner detectors on their own. So this idea is partly foundational history and partly still-living engineering: hand-crafted features remain valuable when data is scarce, when you need speed and transparency, or for geometric jobs like aligning two photos, where learned features offer no clear advantage.
Run an edge detector on a portrait and the photo dissolves into white lines on black: the outline of the face, the rim of the glasses, the part in the hair — and nothing of the smooth, even-toned cheeks, because flat regions have no edges to find.
Edges live where brightness changes sharply; flat regions vanish. They are the raw landmarks recognition builds on.
A common misconception is that deep learning made hand-crafted features obsolete. In practice, the first layers of trained networks reinvent edge and corner detectors anyway, and explicit features still win for low-data, low-power, or geometric tasks like image alignment. The two approaches are cousins, not rivals.