Computer Vision

Introduction to Computer Vision

For a human, looking at a photograph and identifying a “Cat” is an instantaneous, effortless subconscious process. For a computer, a photograph is nothing more than a massive, chaotic grid of millions of numbers (pixels) representing Red, Green, and Blue light intensities.

If a cat moves slightly to the left, is hidden in shadow, or is upside down, the numerical grid completely changes. For decades, traditional software engineers tried to write explicit IF/THEN rules to detect edges and shapes, but these rigid rules failed in the chaotic real world.

Computer Vision (CV) is the field of Artificial Intelligence dedicated to training neural networks to derive high-level semantic meaning from digital images and videos. It is the technology that allows self-driving cars to see pedestrians, and radiologists to detect microscopic tumors.

The Breakthrough: Convolutional Neural Networks (CNNs)

The modern era of Computer Vision began in 2012 with a neural network architecture called AlexNet, which absolutely decimated the competition in the ImageNet image recognition challenge. AlexNet popularized the Convolutional Neural Network (CNN).

CNNs revolutionized image processing by mimicking the human visual cortex. They do not look at the entire image all at once. Instead, they use mathematical “Filters” (Convolutions) that slide across the image, looking for patterns.

CNNs operate in deep, hierarchical layers:

1. Layer 1 (Low-Level): The network scans the pixels and only learns to identify simple shapes: straight lines, vertical edges, and basic colors.
2. Layer 2 (Mid-Level): It combines the straight lines from Layer 1 to identify complex shapes: circles, squares, and curves.
3. Layer 3 (High-Level): It combines the curves from Layer 2 to identify specific semantic concepts: a wheel, an eye, or a furry ear.
4. Final Output: It synthesizes all the high-level features and confidently classifies the image as a “Car” or a “Cat”.

Core Computer Vision Tasks

Image Classification (saying “This is a cat”) is only the most basic form of Computer Vision. Enterprise systems rely on much more complex tasks.

• Object Detection: The AI does not just classify the image; it draws a “Bounding Box” around specific objects. (e.g., A self-driving car drawing a box around 3 different pedestrians and 2 traffic lights in real-time).
• Semantic Segmentation: The AI classifies every single pixel in the image. (e.g., In satellite imagery, the AI colors every pixel of forest green, every pixel of water blue, and every pixel of a road grey, allowing for exact area calculations).
• Facial Recognition: The AI maps the precise geometric distances between a user’s eyes, nose, and mouth, converting the face into a mathematical vector to unlock a smartphone or identify a suspect in a crowd.

The Shift to Vision Transformers (ViTs)

While CNNs dominated the 2010s, the 2020s saw a massive architectural shift.

Google researchers realized that the Transformer architecture (which powers text LLMs like GPT-4) could be adapted for images. They created the Vision Transformer (ViT). Instead of sliding filters across an image like a CNN, a ViT takes an image, slices it into 16x16 pixel “patches,” and treats each patch exactly like a “word” in a sentence. It uses the Attention Mechanism to figure out how every patch of the image relates to every other patch.

Because Transformers are easier to scale on massive GPU clusters than CNNs, ViTs (and hybrid models) have largely taken over the absolute frontier of Computer Vision, serving as the visual “eyes” for natively Multi-Modal models like GPT-4o and Gemini.

Conclusion

Computer Vision is the technology that finally allowed Artificial Intelligence to escape the confines of text databases and interact safely with the physical world. By mathematically extracting meaning from light and pixels, CV serves as the foundational sensory input for robotics, autonomous vehicles, automated manufacturing, and modern medical diagnostics.