AI Computer Vision

What Computer Vision Actually Does

Computer vision is the branch of AI concerned with getting a system to extract meaning from visual input -- photographs, video frames, medical scans, satellite imagery, handwriting. A person can glance at a picture and instantly know it shows a bicycle leaning against a fence; a computer vision system arrives at an equivalent output, such as a label, a set of coordinates, or a yes/no decision, entirely by processing numbers. It never 'sees' the way a person does -- everything it works with starts as data.

An Image Is Really a Grid of Numbers

Every digital image is stored as a grid of pixels, and every pixel is nothing more than a number, or a small set of numbers. A grayscale image is a single 2D grid where each cell holds a brightness value, usually somewhere from 0 (pure black) to 255 (pure white). A color image adds two more grids on top of that -- one for red intensity, one for green, one for blue -- so one pixel might be stored as three numbers like (34, 139, 34), which happens to render as a shade of forest green. A modest 200 by 200 color photo is really an array of 120,000 numbers (200 x 200 x 3 channels) with no built-in concept of 'tree' or 'sky' anywhere in it.

# A tiny 4x4 grayscale image, values from 0 (black) to 255 (white)
image = [
  [ 10,  12, 200, 210],
  [  8,  15, 205, 198],
  [  9,  11, 190, 202],
  [ 12,  14, 195, 205]
]

# To a model this is just a 4x4 array of integers. But notice the
# left two columns are all low numbers (dark) and the right two
# columns are all high numbers (light) -- that sharp jump between
# columns 2 and 3 is what a model would learn to recognize as an edge.

A model never receives a label like 'edge' or 'fur texture' -- it only receives grids of numbers like the one above. Early layers of a vision model learn to notice very simple regularities in those raw pixel values: a sudden jump from dark to light, a patch of repeating texture, a blob of similar color. Deeper layers combine those simple patterns into more complex ones -- edges combine into shapes, shapes combine into parts such as an eye, a wheel, or a window, and parts combine into whole objects such as a face or a car. This step-by-step build-up from raw numbers to recognizable concepts is the essence of how a vision model learns.

Common Computer Vision Tasks

Getting an Image Ready for a Model

1. Resizing every image to the same fixed width and height, since most models expect a consistent input shape.
2. Normalizing pixel values, often rescaling the 0-255 range down to 0-1 or centering it around zero, which helps training behave more smoothly.
3. Augmenting the training set with flipped, rotated, cropped, or recolored copies of existing images so the model doesn't overfit to one exact presentation.
4. Making sure the channel order (such as red-green-blue versus blue-green-red) matches what the model expects, since a mismatch silently feeds it the wrong colors.

Even with careful preprocessing, computer vision stays genuinely difficult. The same object looks different depending on lighting, angle, distance, and partial occlusion -- a cat photographed from above, in shadow, half-hidden behind a couch, is still 'a cat' to a human but represents a very different grid of numbers than a well-lit frontal photo. Effective models have to learn to see past all that variation to the underlying pattern.