CS5720 - Week 10
Slide 188 of 200

U-Net Architecture for Segmentation

U-Net Architecture

Encoder
↓
Bridge
Decoder
↑
The "U" Shape:
Contracting path (encoder) + Expansive path (decoder) + Skip connections
💡 Key Innovation
Skip connections preserve fine-grained details lost during downsampling, enabling precise pixel-level segmentation.

Key Features

  • 🔽 Contracting Path (Encoder)
    Captures context through repeated convolutions and pooling
  • 🔼 Expansive Path (Decoder)
    Enables precise localization through upsampling
  • ➡️ Skip Connections
    Combines high-resolution features with upsampled output
  • 🎯 Weighted Loss Function
    Handles class imbalance and boundary emphasis
Why U-Net Works:
Perfect balance between what (semantic information from encoder) and where (spatial information from skip connections).

U-Net Applications

🏥
Medical Imaging
Tumor segmentation, organ delineation, cell counting
🛰️
Satellite Imagery
Land cover classification, urban planning, agriculture
🚗
Autonomous Driving
Road segmentation, lane detection, obstacle identification
🏭
Industrial Inspection
Defect detection, quality control, surface analysis
Click any application to explore detailed use cases and implementations!
← Previous Next →
Prepared by Dr. Gorkem Kar