Pixels, Channels, and Resolution

Pixels

Part of: Computer Vision Basics

A 12-megapixel phone photo is roughly 36 million numbers. Send that straight to a vision API and two things happen: the model quietly shrinks it before looking, and your bill goes up for pixels the model never really used. Lesson 1 told you an image is a grid of numbers. This lesson is about how many numbers, and why resolution is the single biggest lever on both what a model can see and what it costs you. What it is Resolution is just the image's pixel dimensions: width times height. A photo that is 1920 × 1080 holds about two million pixels per channel, and three channels of color, so roughly six million numbers . Double the width and height and you don't double the count, you quadruple it, because area grows with the square of the side. That quadratic scaling is the whole story of why big images are expensive. The numbers themselves are unchanged from lesson 1: each pixel is 0 to 255 per channel , three channels for RGB. What changes here is the count , and the count is what drives memory, speed, and money. How it works Most vision systems don't accept your image at full size. They resize it to a fixed input, a CNN might want exactly 224 × 224; a multimodal API caps the longest

Challenge: The Resize Budget