All case studies
ShippedEducation · Computer Vision2025Project-Based Learning — First Place
DIP Learning Simulator
Interactive Digital Image Processing simulator that turns theory-heavy labs into a live, chainable pipeline.
Role · Sole engineer
Overview
A full-stack Digital Image Processing learning platform spanning five modules and nineteen interactive tools, built to replace static slides with something students can actually touch. A Python/FastAPI backend runs the algorithms; a React and Three.js frontend visualizes them in real time.
Problem
Image processing courses lean on theory and disconnected lab scripts. Students rarely build intuition for how filters compose, where they break, and what the parameters actually do — because the feedback loop is too slow.
Solution
The core DIP algorithms — enhancement, restoration, segmentation, filtering, feature extraction, and frequency-domain transforms — implemented end to end over OpenCV and NumPy and exposed as parameter-tunable tools rather than static demos. The platform then goes past running the algorithm: automated code explanations, PDF lab-report generation, and session recording all target the gap between executing an operation and understanding it.
Architecture
A Python/FastAPI backend owns the image-processing work and exposes each operation over HTTP; a React frontend drives the parameters and uses Three.js to visualize results in real time, so a parameter change is felt immediately rather than waiting on a page cycle.
Tech stack
- Python
- FastAPI
- React
- Three.js
- OpenCV
- NumPy
Engineering decisions
- Operations modeled as pure functions, not classes, to keep the surface area trivial to extend.
- Every operation exposed as a tunable tool rather than a fixed demo, so the parameters are the lesson.
- Intermediate results cached per pipeline step to keep parameter tweaks responsive.
Challenges
- Frequency-domain visualizations were unintuitive for first-time users; addressed by always pairing magnitude/phase output with the inverse-transformed result.
Performance
- Sub-second interactive feedback on 1024×1024 images for the full operation catalog on commodity hardware.
Lessons learned
- Pedagogical tools succeed or fail on latency, not feature count.
- A small, well-typed plugin contract beats a clever framework.
Future improvements
- Notebook export for any assembled pipeline.
- Shareable pipeline URLs for instructors.
Links
Links coming soon.