The beauty is that nobody programmes these filters by hand. The network learns them during training, starting from random noise and gradually sharpening into useful detectors.

Classification, Detection, and Segmentation

Computer vision tackles three progressively harder tasks:

| Task | Question it answers | Example | |------|-------------------|---------| | Image classification | What is in this image? | "This X-ray shows pneumonia." | | Object detection | What is in this image and where? | Drawing boxes around every pedestrian in a street scene. | | Semantic segmentation | Which pixels belong to which object? | Colouring every pixel of a road, pavement, car, and sky differently. |

Self-driving cars need all three simultaneously - classifying objects, locating them precisely, and understanding the full scene pixel by pixel.

Each task requires progressively more computational power and training data. Classification was largely solved by 2015; real-time segmentation on video remains an active area of research today.

Real-World Applications

Computer vision is already embedded in industries you might not expect:

• Tesla Autopilot uses eight cameras and vision-based AI to detect lanes, traffic lights, and obstacles in real time - processing millions of frames per journey.
• Medical imaging - AI models now match or exceed radiologists at spotting early-stage breast cancer in mammograms, sometimes catching what six human experts missed.
• Quality control - factories use vision systems to inspect thousands of products per minute, catching defects far too subtle or fast for human inspectors.
• Agriculture - drones with computer vision identify diseased crops across vast fields, enabling targeted treatment that reduces pesticide use by up to 90%.
• Retail - Amazon Go stores use computer vision to track which products shoppers pick up, enabling checkout-free shopping.

Ethical Concerns

Computer vision is powerful, but it raises serious questions that society is still grappling with:

• Surveillance - facial recognition enables mass tracking of citizens. Several cities, including San Francisco and parts of the EU, have banned or restricted its use by police.
• Bias - landmark studies by Joy Buolamwini at MIT showed that commercial facial recognition systems were significantly less accurate for darker-skinned faces and women, because training data has historically over-represented lighter-skinned males.
• Consent - should your face be scanned without your knowledge in shops, airports, or public spaces? Many countries are still drafting legislation to address this.
• Deepfakes - AI-generated fake images and videos can spread misinformation and damage reputations, making visual evidence less trustworthy.

Key Takeaways

• Images are grids of pixel values across colour channels - computers see numbers, not pictures.
• CNNs learn to extract features automatically through training, starting from edges and building up to complex objects.
• Classification, detection, and segmentation represent increasing levels of visual understanding.
• Computer vision drives breakthroughs from healthcare diagnostics to autonomous vehicles and precision agriculture.
• Bias in training data and surveillance concerns demand careful, ethical deployment - technology alone is never enough without responsible governance.