All research
Detection accuracy
84%
YOLOv5 variants compared
5

The problem

Aquaculture is one of India's key economic sectors, but farms lack automated ways to find dead or diseased fish. Spotting them by hand across large ponds is slow and error-prone, and one outbreak can devastate a pond within days.

Approach

  • Built a custom fish detection dataset from photos taken in controlled conditions, with Google Images added for test diversity.
  • Trained and compared five YOLOv5 variants (nano, small, medium, large and extra-large) to balance speed and accuracy for aerial detection. I implemented the medium and large variants.
  • Designed for drones: a camera flies over the ponds and images are processed in real time to pinpoint dead fish.
  • The system outputs bounding boxes with confidence scores so farm operators can find and remove diseased fish quickly.
  • Training ran on Google Colab GPUs.

Results

  • 84% detection accuracy for dead and diseased fish in aerial drone imagery.
  • The large and medium variants gave the best balance of accuracy and inference speed for real-time use.
  • A low-cost, software-only approach that works with any commercial drone that carries a camera.