all projects
Android · BTech Final Year Project · 2026

Retina.

An Android app that reads text, money and objects aloud for visually impaired users.

Retina runs four models on the phone camera feed and speaks the results. Everything runs on the device. The paper was presented at ICFISN 2026.

  • Kotlin
  • CameraX
  • ML Kit
  • TensorFlow Lite
  • YOLOv8n
  • MiDaS
fig. 01
Four reading modes, spoken aloud

What I built

  • My contribution: trained the YOLOv8 currency model, built the CameraX real-time frame pipeline, and integrated on-device inference with vibration feedback.
  • Text reader using ML Kit OCR, for labels, mail and signs.
  • Currency reader: a YOLOv8n model I trained on 6 Indian note denominations, with a running total.
  • Object detection with EfficientDet Lite0, announced as left, centre or right.
  • Obstacle warnings from MiDaS depth estimation, with faster beeps as things get closer.
  • No on-screen buttons. Swipes and taps control everything, with vibration feedback.

Hard parts

CHALLENGE // 01

Camera frames arrive faster than the models run

Problem
Queued frames filled memory and caused freezes, so warnings arrived late.
Fix
Used CameraX's keep-only-latest strategy so the model always gets the newest frame.
Learned
For live assistance a fresh answer beats a complete one.
CHALLENGE // 02

Too much speech

Problem
Detections fire every frame, and reading them all out drowned the user in audio.
Fix
Skip text that is nearly identical to what was just said (Levenshtein distance). Collision warnings still interrupt.