Dhruvraj Singh Shekhawat

Projects Bay 04

Exit 1Test track

Semantic segmentation with U-Net

Aerial imagery, Smart India Hackathon project

A deep-learning model for semantic segmentation of aerial imagery, built with Keras and TensorFlow and aimed at on-device use.

Status
Hackathon project
Categories
AI / ML, Computer vision
Built with
Python, Keras, TensorFlow
U-Net: an encoder downsamples, a bottleneck holds the coarsest features, and a decoder upsamples with skip connections to a per-pixel class map. Encoder Decoder skip connections mask
Concept schematic drawn for this page, not a screenshot.

Problem

Aerial images need every pixel labelled, and the model was meant to run on the device rather than a server.

Solution

A U-Net built in Keras and TensorFlow: an encoder that downsamples to capture context, a decoder that upsamples back to full resolution, and skip connections that carry fine detail across.

Architecture

  1. EncoderConvolution and downsampling
  2. BottleneckLowest-resolution features
  3. DecoderUpsampling with skip connections
  4. OutputPer-pixel class map

Dataset, metrics and the repository link have not been added yet.

Sources: Previous site projects page; GitHub profile README

Comments

Questions about the design or ideas for the next iteration are welcome.

1000 characters left. Plain text; links are allowed but limited.

Loading comments…