Home▸Machine Learning & Neural Networks▸U-Net: Encoder-Decoder Architecture for Image Segmentation

🧩 U-Net: Pixel-Perfect Image Segmentation

Explore U-Net, the encoder-decoder neural network that classifies every pixel in an image, powering medical scans, satellite maps, self-driving cars and image generators.

Machine Learning & Neural Networks3DModerate60 FPS

The simulation visualises image data flowing down through U-Net's contracting encoder and back up through its expanding decoder, showing the skip connections carrying detailed feature maps across the U in 3D.

🔬 What It Demonstrates

🎮 How to Use

Pick a sample input image and press play to animate the forward pass layer by layer, or drag the speed slider and scrub manually to inspect how resolution and feature depth change at each stage.

💡 Did You Know?

The original 2015 U-Net paper won the ISBI cell tracking challenge by a large margin using only about 30 annotated training images, an astonishingly small dataset by deep learning standards.

⚙ Under the hood

Watch an image compress through a contracting encoder and expand back through a decoder, with skip connections restoring fine spatial detail for pixel-perfect segmentation.

unetimage segmentationencoder decodermedical imagingmachine-learningcomputer vision

3D · Three.js / WebGL renderer · 60 FPS target · runs fully client-side, no install