Images as Data: Pixels, Resolution, Features and Convolution

Hello students, welcome to Kwickprep. To you, a photo shows a friend smiling. To a computer, the same photo is just a table of numbers. So how can AI find a face in numbers? Today we learn pixels and resolution, greyscale and RGB values, image features, and convolution, step by step.

Let us begin with the building block of every image. The word pixel is short for picture element. A pixel is the smallest square of colour in a digital image. Pixels are arranged in rows and columns, like squares on a chessboard. Each pixel stores a number, or a small set of numbers, that describes its colour.

Resolution is the number of pixels in an image, written as width by height. Six forty by four eighty gives about three lakh pixels, which looks blurry today. Twelve eighty by seven twenty is called HD, with about nine lakh pixels. Nineteen twenty by ten eighty is Full HD, with about twenty lakh pixels. Thirty eight forty by twenty one sixty is called 4K, with about eighty three lakh pixels.

Resolution affects both quality and size. More pixels capture finer detail, so the image looks sharper. But more pixels also make the file bigger and use more memory. If you zoom into a low resolution image, you see blocky squares, called pixelation. That is why AI models often shrink images to a fixed small size before processing them.

In a greyscale image, each pixel holds one number showing its brightness. Zero means black, with no light at all. One hundred twenty eight is a middle grey. Two hundred fifty five means white, the brightest value. These values run from zero to two hundred fifty five because each pixel uses eight bits, which give two hundred fifty six possible values.

A colour pixel stores three numbers, for red, green and blue light, called RGB. Full red with no green or blue gives red. Full green alone gives green. Full blue alone gives blue. Pause and predict: what do full red and full green make? They make yellow, and all three at full give white.

Each set of values is called a channel, like a separate layer for red. A greyscale image has one channel, while a colour image has three channels. So the total numbers in a colour image are width times height times three. A small one hundred by one hundred colour image already holds thirty thousand numbers. This is why images are heavy data for AI.

A feature is a small, useful pattern in an image that helps recognise objects. An edge is a place where pixel values change suddenly, like the outline of a face. A corner is where two edges meet, like the corner of a book. A blob is a patch of similar colour or brightness, like an eye or a mango. Texture is a repeated pattern, like the weave of cloth or blades of grass.

Why do we care about features? A single pixel tells very little, but a pattern of edges can reveal a shape. Good features stay stable even when the lighting or position changes a bit. In early computer vision, humans wrote rules to find these features. Today, deep learning models learn the best features by themselves, using a tool called convolution.

This next idea is the heart of computer vision. Convolution is a way to find features using a small grid of numbers called a kernel, or filter. We place the kernel on the top-left corner of the image, then slide it one step at a time. At each position, we multiply each kernel number with the pixel under it, and add all the products. Every sum becomes one cell of a new grid, called a feature map.

Let us try a tiny example with a three by three image and a two by two kernel. Row one of the image is nine, nine, zero, and the kernel's first row is one and minus one. Row two is again nine, nine, zero, and the kernel's second row is also one and minus one. Row three is nine, nine, zero, and there is a sharp edge where nine drops to zero. Real images use values up to two fifty five, but small numbers keep the maths easy.

Now watch the kernel slide across the image. At the top-left, it covers four nines, so nine minus nine plus nine minus nine gives zero. It slides right, covering nine and zero twice, so nine minus zero plus nine minus zero gives eighteen. It moves down to the bottom-left and covers four nines again, giving zero. At the bottom-right, it covers nine and zero again, giving eighteen.

Here is the finished feature map, two by two. The first row is zero and eighteen. The second row is also zero and eighteen. Zero means no edge, and eighteen means a strong edge, exactly where nine drops to zero. So this kernel is a vertical edge detector.

Let us revise what we learned today. An image is a grid of pixels, and resolution tells how many there are. Greyscale pixels go from zero to two fifty five, and colour pixels store red, green and blue values. Features are useful patterns like edges, corners, blobs and textures. Convolution slides a kernel over the image, multiplying and adding at every step. The feature map shows where that pattern appears. Try the same kernel on an image with rows of zero, zero, nine.