The visual revolution

How deep learning rewrote the rules of image recognition

The Visual Revolution

Start here: vision & CNNs

Computer vision is machine learning on images and video. You take pixel data (sometimes with sensor metadata) and generate outputs a product can use, class labels, bounding boxes, segmentation masks, depth maps, poses, object tracks, captions, or control signals.

Computer vision problems are commonly framed as:

Classification, predicting discrete categories (e.g. cat vs. dog, defect vs. no-defect)
Regression, predicting continuous values like coordinates, depth, or motion vectors
Vision + language tasks, generating or understanding text from images, such as image captioning or visual question answering. These are usually framed as conditional language modeling, a sequence-generation problem conditioned on an image.

In production, that almost always means coping with noise, lighting, motion blur, occlusion, and data that doesn’t match your training set. Your job as an engineer is to design and train models that stay correct and robust enough when the world refuses to look like the benchmark.

Introduction to vision as a human and machine capability

For humans, sight is one of the main ways we understand the world, reading faces, streets, and traffic lights, often without thinking hard about each step. This lesson is about the same idea for machines: using raw images or video so computers can see (in a useful, engineered sense). Deep learning is the usual tool today for turning pixels into predictions.

Human vision

Eyes capture light and the brain interprets scenes, objects, depth, and motion.
Humans perceive meaning directly.
Vision is continuous and semantic.
Humans reason using memory and context.
Humans generalize naturally across new scenes.
Output: perception and action.

Machine vision

Cameras capture images as pixel grids represented numerically as /matrices.
Models process numerical pixel values to learn patterns and representations.
Neural networks transform pixel tensors into embeddings, predictions, and actions.
Models learn statistical patterns from training data.
Output: classification, detection, segmentation, captioning, tracking, or control signals.

Concretely, an image enters a model as a numeric tensor whose shape depends on how many channels it carries:

Grayscale image → 2D matrix of shape (H × W), one intensity value per pixel.
RGB image → 3D tensor of shape (H × W × 3), where each pixel contains Red, Green, and Blue intensity values.

Same patch, two tensor shapes

Grayscale · (H × W) — one value per pixel

3014090200255601280175

shape (3 × 3) = 9 numbers

RGB · (H × W × 3) — three values per pixel