Ch5. Computer Vision — How Computers See the World

What Is Computer Vision?

Computer Vision is the AI field focused on extracting meaningful information from images and video. It mimics the human visual system — and in specific tasks, surpasses it.

Applications: autonomous vehicles, medical image diagnosis, facial recognition, quality inspection, augmented reality, satellite imagery analysis.

---

CNN: The Architecture Behind Modern Vision

Convolutional Neural Networks (CNNs) are the backbone of modern computer vision.

Convolution: Detecting Patterns

The core operation that finds patterns (edges, textures, shapes) in images:

Input image: 228×228×3 (height×width×RGB channels)
Filter (kernel): 3×3 learnable weights
Output: a feature map showing how strongly each position activates

Key insight: The same filter slides across the entire image (parameter sharing). A cat-ear filter detects ears wherever they appear in the image.

Pooling: Reducing Dimensions

Pooling reduces the feature map size while retaining the most important information:

• MaxPooling: Keep the maximum value in each 2×2 region
• Creates translation invariance: small shifts in the image produce the same features

CNN Layer Hierarchy

Input image
→ Conv (edge detection) → Conv (texture) → Conv (patterns) → Conv (high-level features)
→ Flatten → Fully Connected → Output

Lower layers: edges, lines (low-level) Higher layers: eyes, faces, objects (high-level)

---

Major Computer Vision Tasks

1. Image Classification

Assign a single category label to an entire image.

• Example: “Is this photo a cat or a dog?”
• Key models: AlexNet (2012 breakthrough), VGG, ResNet, EfficientNet

2. Object Detection

Simultaneously locate (bounding box) and classify multiple objects in an image.

• Example: Autonomous car detecting pedestrians, traffic lights, lane markings
• Key models: YOLO series, Faster R-CNN, SSD

Model	Speed	Accuracy	Use Case
YOLO	Very fast	Good	Real-time detection
Faster R-CNN	Slower	Very high	Precision analysis

3. Semantic Segmentation

Classify every pixel with a category label.

• Example: Autonomous driving — every pixel labeled as “road,” “pedestrian,” or “vehicle”
• Key models: U-Net (medical imaging), DeepLab

4. Instance Segmentation

Semantic segmentation + distinguishing individual object instances.

• Example: Identify each specific person in a crowd separately

---

Computer Vision in Medicine

Skin cancer diagnosis: A Stanford team trained a CNN to classify skin lesions at dermatologist level (2017, Nature).

Retinal disease: Google DeepMind’s system diagnoses 50+ retinal conditions at specialist-level accuracy.

Radiology: AI systems detect pneumonia patterns in chest X-rays, helping prioritize urgent cases.

---

Key Concept Cards

CNN ★★★★★ : Convolution→Pooling→Classification hierarchy. Learns hierarchical image features via parameter sharing. The standard architecture for visual AI.

YOLO ★★★★☆ : “You Only Look Once.” Single forward pass detects all objects simultaneously. Essential for real-time applications: autonomous vehicles, CCTV analysis.

Transfer Learning ★★★★★ : Fine-tune a model pre-trained on large data for your specific task. Industry standard — rarely train from scratch.

---

Practice Quiz

Q1. Why is YOLO faster than Faster R-CNN for real-time applications?

YOLO divides the image into a grid and predicts all bounding boxes and class probabilities in a single neural network pass. Faster R-CNN uses a two-stage pipeline: first generate region proposals (RPN), then classify each region. YOLO’s single-pass approach is much faster, though slightly less precise on small objects.

Q2. Why is U-Net particularly suited to medical image segmentation?

U-Net’s encoder-decoder structure with skip connections combines low-resolution contextual information (from deep layers) with high-resolution spatial detail (from early layers). This allows precise pixel-level segmentation even with limited training data — crucial in medical imaging where labeled data is scarce and expensive.